A hallmark of CNS organization is the highly precise pattern of connectivity between neurons. The location of a synapse on a target cell and its distinctive structural and functional properties are key factors in determining the flow of information in a neural circuit.

The nervous system uses many mechanisms to accomplish the formidable task of connecting neurons with their appropriate target cells (reviewed in Ref. 1). The organization of many brain regions into distinct anatomical layers (laminae) is one key structural feature that aids wiring specificity. The targeting of processes to specific laminae, which we refer to as laminar specificity, limits the range of target cells or surfaces available for synapse formation. Within a particular lamina, processes can synapse with a specific cell type that resides in that layer, such as a principal cell or an interneuron (cellular specificity). Neurons can also form synapses onto a specific subcellular domain of a target cell, such as the axon initial segment or dendritic compartment (subcellular specificity). In the final step of assembling functionally connected circuits, synapses between different types of neurons are differentiated into structurally and functionally distinct synapse types (synaptic diversity). These different levels of specificity are closely related. For example, dendrites from principal neurons in brain regions such as the hippocampus or the cortex span multiple laminae and form a different type of synapse with a distinct presynaptic partner in each lamina. Depending on the viewpoint, this could be considered as an example of laminar specificity, subcellular specificity or synaptic diversity.

The molecular mechanisms that regulate wiring specificity and synaptic diversity in the vertebrate brain are only beginning to be understood. Classic work from Langley2 and Sperry3 on the regeneration of nerve fibres in the mature nervous system indicated that neurons can rewire with remarkable specificity. Their work suggested the presence of “individual identification tags” on cells and fibres that would allow neurons to distinguish one another and selectively connect to target cells by “specific chemical affinities” (Ref. 3). In recent years, substantial progress has been made in identifying and characterizing molecularly diverse cell surface protein families, which may to some extent act as the surface tags envisioned by Langley and Sperry. Advances in genomics and proteomics are enabling the analysis of cell type-specific repertoires of surface proteins and the systematic, large-scale mapping of their extracellular interaction networks. Despite these advances, determining how this molecular diversity contributes to the encoding of wiring specificity and synaptic diversity remains a considerable challenge.

Several requirements can be envisioned for cell surface proteins that regulate wiring specificity and synaptic diversity. Such proteins should be expressed in distinct populations of neurons, be capable of interacting in trans with their binding partners, and provide enough molecular diversity to confer cell type- and synapse type-specific identities. In this Review, we start by introducing several superfamilies of cell surface proteins that meet these requirements: the cadherins, neurexins, leucine-rich repeat (LRR) proteins and immunoglobulin (Ig) proteins. The molecular diversity of these protein families arises either through the large size of the underlying gene family or through the alternative splicing of a more-limited number of genes. We review the evidence for cell type-specific expression of these proteins and discuss to what extent current in vivo evidence supports a role for them in wiring specificity and synaptic diversity. Other cell surface proteins exhibiting cell type-specific expression patterns have been implicated in the specification of connectivity, and a brief, albeit incomplete, overview of additional molecular players in this process is provided in Box 1. Note that, owing to space limitations, we do not discuss the cellular and molecular mechanisms that guide axons towards their target cells, nor do we provide an overview of all the cellular and molecular mechanisms that are known to contribute to synaptic specificity, as these have been extensively reviewed elsewhere1,4,5,6. Finally, this Review focuses on the rodent CNS, and we refer the reader to several excellent recent reviews that discuss synaptic specificity in invertebrate systems7,8,9.

Cell surface protein families

Classic cadherins and protocadherins. The cadherin superfamily comprises more than 100 transmembrane glycoproteins that can be grouped into several subfamilies, of which the classic cadherins and protocadherins have been the most studied in relation to connectivity10,11,12,13. Classic cadherins contain five extracellular cadherin (EC) repeats and a conserved intracellular domain that contains catenin-binding sites (Fig. 1A). Approximately 20 classic cadherins exist, and these are further classified as type I or type II cadherins based on amino acid sequence identity in their first EC repeat11. Classic cadherin trans interactions are mediated by the EC1 domain (Fig. 1A) and are preferentially homophilic, but they can also be heterophilic within each class; thus, type I cadherins can interact with other type I cadherins, and type II cadherins can interact with other type II cadherins14,15,16,17,18.

Figure 1: Domain organization and interactions of classic cadherins, protocadherins and neurexins.
figure 1

A | Classic cadherins contain five extracellular cadherin (EC) repeats and a largely constant intracellular domain containing p120-catenin- and β-catenin-binding sites. The EC1 domains of trans-interacting cadherins mediate recognition specificity. B | Protocadherins contain six EC repeats and their cytoplasmic tails lack canonical catenin-binding sites. Studies in non-neuronal cells suggest that protocadherin trans interactions are homophilic and mediated by their EC2 and EC3 domains. The protocadherin recognition unit is a multimer, possibly a tetramer, as shown here. Ca | Neurexins can exist as long (α) and short (β) forms. The α-neurexin (NRXNα) proteins contain six extracellular laminin–neurexin–sex-hormone-binding globulin (LNS) domains, interspersed with three epidermal growth factor (EGF)-like domains, and contain a carbohydrate-attachment sequence near the plasma membrane. The NRXNα extracellular domain contains up to six splice sites (SS1–SS6). β-neurexin (NRXNβ) proteins contain a short amino-terminal NRXNβ-specific leader sequence, a single LNS domain and a carbohydrate-attachment sequence. Their extracellular domain contains two splice sites (SS4 and SS5). The carboxyl terminus of all neurexins contains a PDZ (postsynaptic density 95, Discs large 1 and zonula occludens 1)-binding domain. Cb–Cf | Presynaptic neurexins interact heterophilically in trans with multiple (postsynaptic) binding partners, which can interact specifically with NRXNα or NRXNβ (Cb,Cc), or with both NRXNα and NRXNβ (note that only the NRXNα interactions are shown) (Cd–Cf). Calsyntenin (CLSTN) and neurexophilin (NXPH) are NRXNα-specific binding partners. CLSTN interacts with the NRXNα LNS5–EGF3–LNS6 domains (Cb). NXPH interacts with the LNS2 domain of NRXNα (Cb). A splice variant of neuroligin 1 that contains a 'B' splice insert (NLGN1 (B+)) is a NRXNβ-specific binding partner (Cc). Neurexin splicing at SS4 in LNS6, which is common to both NRXNα and NRXNβ, regulates neurexin–ligand interactions. The presence of SS4 (SS4+) modulates the interaction with NLGN1 (B), NLGN2, NLGN3 and NLGN4 (Cd). Neurexins lacking an SS4 insert (SS4) interact with leucine-rich repeat transmembrane neuronal protein (LRRTM), dystroglycan (which can also interact with the LNS2 domain of NRXNα) and latrophilin 1 (LPHN1) (Ce). SS4+ neurexins interact with the secreted ligand cerebellin (CBLN), which forms a complex with the δ2 glutamate receptor (GluD2) (Cf). LRRCT, LRR C-terminal domain; LRRNT, LRR N-terminal domain.

PowerPoint slide

The protocadherins were identified based on their structural similarity to cadherins19 and are subdivided into clustered and non-clustered protocadherins. The role of non-clustered protocadherins in connectivity is not well understood20 and is not discussed here. The clustered protocadherins comprise the α-, β- and γ-clusters, which are arranged in three tandem arrays on a single chromosome that encodes a total of 58 isoforms in mice21,22,23. The α- and γ-protocadherin clusters contain variable exons encoding complete extracellular, transmembrane and proximal intracellular domains. Each exon contains its own promoter and is spliced to three constant exons that encode the common remainder of the cytoplasmic domain24,25. The β-cluster lacks constant exons and instead contains exons that seem to encode complete proteins. Five phylogenetically divergent members from the α- and γ-clusters, which differ from their respective cluster members and are more similar to each other, have been grouped into a separate C-cluster23.

Although structurally similar, clustered protocadherins differ from classic cadherins in several aspects, with direct implications for their role in connectivity. Protocadherins contain six EC repeats instead of five, and their cytoplasmic tails lack canonical catenin-binding sites (Fig. 1B). Protocadherin isoforms from the α-, β- and γ-gene clusters engage in highly specific homophilic trans interactions that are mediated by the EC2 and EC3 domains, instead of by the EC1 domain26. Protocadherins can associate in cis and form multimers (Fig. 1B), which have different trans binding specificities than individual isoforms alone and display remarkable recognition specificity26,27. For example, cells expressing a combination of five different protocadherin isoforms will co-aggregate with a second set of cells expressing the same combination of isoforms. However, the inclusion of a single mismatched isoform in the second set of cells is sufficient to prevent co-aggregation27. Thus, co-expression of multiple protocadherin isoforms greatly diversifies recognition specificity, and protocadherin multimers consisting of random isoform combinations would have enormous potential for unique adhesive interactions. However, it should be noted that these aggregation experiments have been performed in heterologous cells and that much remains uncertain about the exact nature of protocadherin recognition27, as well as about the composition of putative protocadherin multimers in neurons. In addition, protocadherins can form cis complexes with other surface proteins, including classic cadherins28 and the GABAA receptor (GABAAR)29, which may affect protocadherin trans interactions.

Neurexins. In contrast to the classic cadherins and protocadherins, the neurexin family consists of only three genes (Nrxn1Nrxn3), each of which has two alternative promoters that allow the generation of long (α) and short (β) forms of neurexins. The α-neurexins contain six extracellular laminin–neurexin–sex-hormone-binding globulin (LNS) domains, interspersed with three epidermal growth factor (EGF)-like domains30 (Fig. 1C). The shorter β-neurexins contain a single LNS domain. Neurexins all share the same short intracellular domain but contain a varying number of splice sites in their ectodomains (six for NRXN1α and NRXN3α, five for NRXN2α and two for β-neurexins). Extensive alternative splicing of neurexins was noted soon after their discovery31,32. The theoretical total number of possible neurexin variants ranges in the thousands33,34, but the true extent of neurexin alternative splicing has become clearer with recent deep-sequencing approaches35,36. Comprehensive quantitative analysis of neurexin diversity in the adult mouse cortex has indicated near-exhaustive combinatorial alternative exon use in Nrxn1a and Nrxn2a transcripts, resulting in hundreds of observed variants36. However, Nrxn3a transcript diversity is much lower than the theoretical number of such transcripts, resulting in hundreds, rather than the theoretical thousands, of observed NRXN3α variants.

The alternative splicing of neurexins regulates neurexin–ligand interactions. Neurexins are thought to be predominantly presynaptically localized and to engage in heterophilic trans interactions with multiple, highly diverse binding partners. The first neurexin ligand to be identified was the postsynaptic adhesion molecule neuroligin37. Neuroligin binding to neurexin induces presynaptic differentiation, and neurexin binding to neuroligin induces postsynaptic differentiation38,39,40,41. The neurexin–neuroligin interaction plays a major part in the formation, maturation and function of synapses42,43. Additional neurexin ligands, transmembrane and secreted, have since been identified. Neurexins interact with various transmembrane proteins, namely LRR transmembrane neuronal proteins (LRRTMs)44,45,46, calsyntenin47, dystroglycan48, latrophilin (LPHN)49 and the GABAAR50, as well as with the secreted molecules neurexophilin51,52 and cerebellin53 (Fig. 1C). Of these, LRRTMs, calsyntenins and cerebellins have a capacity to induce presynaptic differentiation similar to that of the neuroligins. Most neurexin ligands interact with the last LNS domain, which is common to both α- and β-neurexins. This LNS domain contains alternative splice site 4 (SS4), which is the best-characterized splice site in terms of regulating neurexin–ligand interactions. For some neurexin ligands, SS4 acts as a molecular switch in determining binding. LRRTM2, LPHN1 and dystroglycan only bind to neurexin that lacks SS4 (Refs 45,46,48,49), whereas cerebellin 1 exclusively binds to neurexin containing a small SS4 insert53 (Fig. 1C). By contrast, for neuroligin, the presence of SS4 in neurexin acts as a modulator that decreases the affinity of the neurexin–neuroligin interaction54. The remaining, largely uncharacterized, splice sites probably also modulate neurexin interactions with postsynaptic ligands. Consistent with this possibility, a recent study using proteomic neurexin isoform profiling showed that LRRTM2 preferentially associates with NRXN1α and NRXN2α over NRXN1β and NRXN2β, and that the presence of the SS6 insert negatively modulates the NRXN1α–LRRTM2 interaction55.

LRR and Ig protein superfamilies. The LRR and Ig domains are major protein–protein interaction motifs that are often found in cell surface molecules. The LRR protein superfamily is considerably larger in vertebrates than in Drosophila melanogaster56. The LRR domain is a versatile protein-interaction domain57, and recent studies indicate that neuronal LRR proteins engage in molecularly diverse trans interactions58 (Fig. 2A). In contrast to LRRTM2, LRRTM4 preferentially interacts with presynaptic heparan sulphate proteoglycans59,60. Fibronectin-like domain-containing leucine-rich transmembrane protein 3 (FLRT3), another postsynaptic LRR protein, interacts with the presynaptic G protein-coupled adhesion receptor LPHN61. The postsynaptic LRR protein netrin-G2 ligand (NGL2; also known as LRRC4) trans-synaptically interacts with the glycosylphosphatidylinositol-anchored protein netrin-G2 (Ref. 62).

Figure 2: Domain organization and interactions of LRR and Ig superfamily proteins.
figure 2

A | The leucine-rich repeat (LRR) domain is made up of individual LRRs, which are usually flanked by an LRR amino-terminal domain (LRRNT) and an LRR carboxy-terminal domain (LRRCT). LRR proteins interact heterophilically in trans with diverse ligands. The LRR domain of the postsynaptic proteins LRR transmembrane neuronal protein 1 (LRRTM1) and LRRTM2 interacts with the sixth laminin–neurexin–sex-hormone-binding globulin (LNS6) domain of presynaptic α-neurexin (NRNXα) or β-neurexin (not shown) that lacks splice site 4 (SS4), whereas LRRTM4 preferentially interacts with the heparan sulphate (HS) chains of presynaptic HS proteoglycans (HSPG), such as glycosylphosphatidylinositol (GPI)-anchored glypican. The postsynaptic protein fibronectin-like domain-containing leucine-rich transmembrane 3 (FLRT3) contains a fibronectin type III (FN3) domain in addition to its LRR domain, which interacts with the olfactomedin domain of presynaptic latrophilin 3 (LPHN3). The LRR domain of postsynaptic netrin-G2 ligand (NGL2) interacts with domain VI (laminin N-terminal domain) of the GPI-anchored presynaptic protein netrin-G262. The extracellular LRR and FN3 domain-containing 1 (ELFN1) protein interacts with presynaptic metabotropic glutamate receptor 7 (mGluR7), although the binding site remains unclear (as indicated by the question mark). The secreted LRR protein Slit homologue 2 (SLIT2) interacts via its second LRR domain with the first immunoglobulin (Ig) domain of its receptor Roundabout homologue 2 (ROBO2). Ba | Ig superfamily proteins of the L1 (also known as neuron–glia cell adhesion molecule (NgCAM)) family, consisting of L1, close homologue of L1 (CHL1), neuronal CAM (NRCAM) and neurofascin (NFASC) have a domain organization consisting of Ig and FN3 domains. The contactin (CNTN), sidekick (SDK), Down syndrome CAM (DSCAM) and DSCAM-like (DSCAML) proteins are closely related to the L1 family proteins and share a similar domain organization. GPI-anchored CNTN signals via CNTN-associated protein (CASPR). Bb | Homophilic interactions of SDKs, CNTNs and DSCAMs have been implicated in the control of laminar specificity in the retina. Bc | A complex consisting of CHL1 and NRCAM interacting in trans with CNTN5 and CASPR4 has been implicated in control of subcellular specificity the spinal cord. CT3, C-terminal cysteine knot domain; LEGF, laminin EGF-like domain.

PowerPoint slide

The Ig domain is among the most-common protein domains in the human genome63. A prominent family of Ig superfamily proteins that has been implicated in connectivity is the L1 (also known as the neuron–glia cell adhesion molecule (NgCAM)) family, consisting of L1, close homologue of L1 (CHL1), neuronal CAM (NRCAM) and neurofascin (Fig. 2B). These proteins can engage in homophilic and heterophilic interactions and have a domain organization consisting of Ig and fibronectin type III (FN3) domains. The L1 family proteins are closely related to the contactin (CNTN), sidekick (SDK), Down syndrome CAM (DSCAM) and DSCAM-like (DSCAML) proteins (Fig. 2B), which bind homophilically but do not seem to interact heterophilically with each other64.

Cell type-specific expression

Before discussing how molecularly diverse cell surface proteins contribute to wiring specificity and synaptic diversity, we first consider evidence for cell type-specific expression of cadherins, neurexins, and LRR and Ig superfamily proteins in the nervous system.

Type I classic cadherins are widely expressed in the nervous system, but type II cadherins are expressed in discrete patterns. Tracing studies showed that a dye injected into specific cortical regions expressing either of the type II cadherins cadherin 6 (Cdh6) or Cdh8 retrogradely labelled, respectively, Cdh6- or Cdh8-expressing thalamic nuclei projecting to these regions, indicating that type II cadherins are expressed in projection-specific patterns65,66. Mapping studies of the expression patterns of multiple classic cadherins in various brain regions corroborated these initial findings and showed that cadherins are expressed in complex, partially overlapping patterns, labelling specific neuronal populations and fibre tracts67,68,69,70,71. Together with the enriched localization of several classic cadherins at synapses72,73,74,75,76, these observations gave rise to the idea that cadherin-based interactions might contribute to cell–cell recognition and wiring specificity76,77,78.

Clustered protocadherins are broadly expressed in the CNS, but single-cell reverse transcription PCR (RT-PCR) analysis has shown that individual Purkinje cells express different combinations of clustered protocadherins79,80,81,82,83,84,85. Their variable single-cell expression and extraordinary potential for encoding unique adhesive interactions suggest that clustered protocadherins might be involved in generating distinct cell surface identities. One role for such identities might be in the recognition of self versus non-self, similarly to the part played by Dscam1 diversity in D. melanogaster86,87. Protocadherins have been detected in synaptic and extrasynaptic membranes, but they are also found in intracellular compartments and it is still unclear to what extent they are present on the neuronal surface22,23,88. Protocadherins remain enigmatic surface molecules and much remains to be discovered about their roles in circuit formation.

The α- and β-forms of Nrxn1–Nrxn3 transcripts are differentially expressed in the nervous system, in discrete but overlapping neuronal populations32. A quantitative comparison of neurexin alternative splicing in different brain regions showed that a complex structure such as the cortex contains a more-varied neurexin repertoire than the relatively less complex cerebellum36. Purified cerebellar granule cells displayed further reduced complexity, indicating that neurexin molecular diversity is linked to cellular diversity36. Alternative splicing of neurexins is regulated in a cell type-specific manner89,90, supporting the idea that different populations of neurons might express distinct repertoires of neurexins. A recent study profiling neurexin isoform expression in single cells (Box 2) that were isolated from neural tissue confirmed that neurexin repertoires are indeed highly cell type-specific and are reproducible for a specific neuron type91. For example, different populations of interneurons located in the same hippocampal lamina and projecting to the same pyramidal target cell population displayed divergent neurexin expression profiles. Different populations of long-range projection neurons converging on the same target cell also displayed distinct neurexin profiles91. Together with the predominantly presynaptic localization of neurexins and the absence of major axon-pathfinding defects in knockout mice deficient in all three α-neurexins92, these findings suggest that cell type-specific neurexin repertoires do not convey target cell specificity, but are more likely to modulate the range of postsynaptic ligands with which these cells interact and, as a result, may contribute to the specification of synaptic properties.

LRR proteins are often expressed in discrete neuronal populations and are generally localized to the postsynaptic surface. For example, Lrrtm4 expression in the hippocampus is restricted to dentate gyrus (DG) granule cells. Postsynaptic LRRTM4 interacts with presynaptic heparan sulphate proteoglycans to regulate excitatory synapse development in these neurons59,60. FLRT3 is expressed in hippocampal granule cells and CA3 pyramidal neurons and regulates granule cell excitatory synapse development via its interaction with presynaptic LPHN3 (Ref. 61). Their discrete expression patterns and diversity in trans interactions suggest that LRR proteins may contribute to the specification of cell type-specific synaptic properties.

A good example of cell type-specific expression of Ig superfamily proteins is found in the retina, in which SDK1 and SDK2, DSCAM, DSCAML and their close relatives CNTN1–CNTN5, are differentially expressed in largely non-overlapping subsets of presynaptic amacrine cells and postsynaptic retinal ganglion cells64,93,94. Different classes of amacrine and retinal ganglion cells extend their processes in distinct synaptic laminae of the inner plexiform layer (IPL), where they form synapses. The discrete expression patterns of SDKs, DSCAMs and CNTNs in matching populations of partner cells, combined with their homophilic binding properties, suggests that these Ig family proteins could regulate lamina-specific targeting of retinal neuron processes.

From surface protein to interactome

As it will become increasingly feasible to elucidate the surface molecule repertoire of specific cell types as technology develops (Box 2), a next step would be to determine the complete cell surface interaction networks, or extracellular interactomes, of these cells. Technical challenges have impeded the identification of extracellular interactions95, but recent advances in genomics, bioinformatics, assay technology and proteomics are enabling novel extracellular interactions to be discovered at a much higher rate than before. Advances in genomics and bioinformatics now allow for systematic surveys of all cell surface proteins encoded in the genomes of model organisms. These analyses have revealed that the expansion of genes encoding CAMs correlates with the increasing cellular complexity of organisms96. For example, the LRR and Ig superfamilies have expanded in the genome of vertebrate organisms compared with invertebrates56,63,96, suggesting that an expanded repertoire of cell surface proteins allows for an increased complexity of connectivity.

Large-scale, systematic analysis of the extracellular interactions between surface proteins has been accelerated by the development of sensitive, high-throughput binding assays. One study developed an enzyme-linked immunosorbent assay (ELISA)-based assay to quantitatively assess thousands of interactions of Dscam1 isoforms in flies97. Similar approaches have since been adopted for LRR and Ig superfamily proteins. These assays make use of secreted 'bait' and 'prey' proteins obtained from the media of transfected cells, and they rely on the clustering of bait proteins, which are immobilized on plates, and the oligomerization of prey proteins to enhance the avidity of low-affinity interactions. Binding is then assessed by an enzymatic reaction98,99,100,101. Large-scale binding assays for nearly all members of zebrafish and D. melanogaster Ig and LRR superfamilies that tested tens of thousands of unique interactions between hundreds of recombinant proteins have yielded close to a hundred novel interactions99,101. Technical problems such as incorrect folding of recombinant bait and prey proteins aside, the outcomes of these experiments are currently limited by the number and type of proteins assayed, suggesting that much of the extracellular interactome remains undiscovered. Another limitation is that the use of recombinant ectodomains in these binding assays does not reflect the complex nature of assemblies of adhesion proteins on the cell surface in vivo. However, scaling-up these assays to eventually test millions of interactions between thousands of proteins could deliver complete extracellular interactomes of specific cell types. These and other technological advances (Box 2) will enable the combination of expression profiles of specific neuronal populations with their extracellular interactomes and connectivity maps102, allowing researchers to start generating models of connectivity, which can then be experimentally tested in model systems.

Surface proteins and wiring specificity

Despite rapid progress in elucidating cell type-specific gene expression profiles and extracellular interactomes, testing how cell surface molecules contribute to specifying connectivity in the vertebrate nervous system is still largely limited to the study of single genes. Screening large numbers of cell surface genes for their roles in connectivity is more feasible in simpler organisms, such as D. melanogaster103. However, improved genetic access to specific cell types increasingly allows testing the consequences of the gain or loss of function of surface proteins in neural connectivity with single-cell resolution in vivo in the vertebrate CNS. Below, we highlight several examples of surface proteins that regulate different aspects of wiring specificity.

Self-avoidance. To achieve the correct wiring specificity, neurons must recognize not only the processes of other neurons but also their own neurites. The propensity of neurites from the same neuron to repel each other, a process termed self-avoidance, is indispensable for proper patterning of the D. melanogaster nervous system87. In D. melanogaster, the extraordinary diversity of the Dscam1 family of Ig superfamily cell surface proteins regulates self-avoidance104.

The diversity of clustered protocadherins and their potential for encoding unique adhesive interactions suggested that protocadherins might have a similar role in vertebrates, which lack DSCAM diversity. Deletion of the γ-protocadherin cluster in retinal starburst amacrine cells (SACs), which have radially symmetric dendrites that rarely cross, resulted in frequent crossing and bundling of dendrites, without affecting their laminar targeting105. This process was cell-autonomous, and replacement of a single γ-protocadherin isoform was sufficient to rescue self-avoidance in SACs (Fig. 3a). However, protocadherin diversity is needed to allow these neurons to distinguish their own dendrites from the dendrites of other neurons, so that dendrites from the same SAC avoid each other but dendrites from different SACs interact freely. A similar role for γ-protocadherins in self-avoidance was found in Purkinje cells105. These results indicate that clustered protocadherins define unique cell surface identities that are important for self-recognition in vertebrates, much like Dscam1 diversity does in flies. However, it remains to be determined to what extent protocadherin-mediated self-avoidance is a general principle, as cortical pyramidal neurons lacking the γ-protocadherin cluster displayed impaired arborization but not bundling or crossing of dendrites106. Furthermore, protocadherins are not the only class of surface proteins mediating self-avoidance in vertebrates. Recent studies have identified a role for multiple EGF-like domain (MEGF) receptors, as well as the transmembrane protein semaphorin 6A and its receptor plexin A2 in SAC self-avoidance107,108, and the secreted LRR protein Slit homologue 2 (SLIT2) and its receptor Roundabout homologue 2 (ROBO2) in Purkinje cell self-avoidance109. Taken together, current in vivo evidence supports a role for protocadherins in self- avoidance, but given the many questions that remain about protocadherin interactions, localization and function, additional roles for the proteins in, for example, synapse development110 can presently not be ruled out.

Figure 3: Protocadherins and classic cadherins in retinal circuit assembly.
figure 3

a | Protocadherins regulate self-avoidance. Retinal starburst amacrine cells (SACs) have radially symmetric dendrites that rarely cross (left-hand panel). Deletion of the entire γ-protocadherin (Pcdhg) gene cluster (encoding 22 isoforms) in SACs results in frequent crossing and even bundling of dendrites (centre panel). Replacement of a single γ-protocadherin isoform in Pcdhg-knockout (KO) SACs is sufficient to rescue self-avoidance in SACs (right-hand panel). b | Type II classic cadherins regulate laminar targeting in the retina. Two distinct types of bipolar cells (BCs), each expressing a different type II cadherin, target different sublaminae in the retinal inner plexiform layer (IPL). Cadherin 8 (CDH8)-expressing 'OFF' type 2 BCs (BC2s; orange) target the outer IPL (the OFF layer), where they synapse onto the dendrites of OFF SACs and the upper branch of the bistratified dendrites of 'ON–OFF' direction-selective ganglion cells (ooDSGCs). CDH9-expressing 'ON' BC5s (blue) target the inner IPL (ON layer), where they synapse onto the dendrites of ON SACs and the lower dendritic branch of ooDSGCs. Loss of Cdh8 in BC2s results in mistargeting of their axonal arbor to the inner IPL, but loss of Cdh8 in BC5s does not affect their axonal targeting. Conversely, loss of Cdh9 in BC5s results in the mistargeting of their axons to the outer IPL, but loss of Cdh9 in BC2s does not affect their axonal targeting. Ectopic expression of CDH9 in BC2s leads to displacement of their axons to the inner IPL, and ectopic expression of CDH8 in BC5s displaces their arbors to the outer IPL. Ectopic expression of CDH9 in BC2s in a Cdh9-KO background still displaces their axons to the inner IPL, and introduction of CDH8 in BC5s in a Cdh8-KO background still displaces their axons to the outer IPL, indicating that type II cadherins instruct laminar targeting via a heterophilic mechanism.

PowerPoint slide

Laminar specificity. Laminar targeting is a key step in establishing specific wiring patterns. Classic experiments using co-cultures demonstrated that type I neural cadherin (N-cadherin) contributes to laminar specificity111,112. Exploiting advances in genetic targeting of specific cell types (Box 2), a recent study demonstrated that type II cadherins regulate laminar targeting in the retina113. Two distinct types of retinal bipolar cells (BCs), each expressing a different type II cadherin, target different synaptic laminae in the retinal IPL. Axons of Cdh8-expressing type 2 BCs (BC2s) target the outer IPL (Fig. 3b), whereas Cdh9-expressing BC5s target the inner IPL. Loss of Cdh8 in BC2s resulted in mistargeting of part of their axonal arbor to the inner IPL. Conversely, loss of Cdh9 in BC5s resulted in partial mistargeting of BC5 axons to the outer IPL (Fig. 3b), suggesting that CDH8 and CDH9 function to bias BC axon choice between inner and outer IPL laminae. Ectopic expression of CDH9 in BC2s redirected their axons to the inner IPL, whereas expression of CDH8 in BC5s redirected their arbors to the outer IPL (Fig. 3b), indicating that CDH8 and CDH9 instruct laminar targeting. Remarkably, introduction of CDH9 in BC2s on a Cdh9−/− background still redirected BC2 axons to the inner IPL (Fig. 3b). These results show that differentially expressed type II cadherins act heterophilically with laminar ligands in instructing the laminar choice of BC axons. Functional synaptic connectivity between BCs and ganglion cells in Cdh8−/− or Cdh9−/− mutants was more-strongly reduced than might be expected from the partial mistargeting defects114, suggesting that, in addition to instructing BC axon laminar choice, CDH8 and CDH9 may promote synapse formation with their target cells.

In addition to the classic cadherins, many other proteins have been implicated in laminar targeting in the IPL. The Ig superfamily proteins SDK1, SDK2, DSCAM and DSCAML, as well as CNTN1–CNTN5, are expressed in discrete, largely non-overlapping subsets of pre- and postsynaptic retinal neurons and are concentrated in specific IPL synaptic laminae64,93,94. Homophilic interactions between such proteins on neurites from matching pre- and postsynaptic partner cells that project to the same IPL lamina have been proposed to promote lamina-specific connectivity. Gain- and loss-of-function studies lend some support to this idea. For example, knockdown of CNTN2 in broad regions of the chick retina results in a more diffuse distribution of presumptive CNTN2-expressing neurites, identified by SDK1, which is co-expressed in a subset of CNTN2-positive neurons. Conversely, overexpression of CNTN2 in groups of retinal cells tends to bias neurites towards CNTN2-positive laminae64. Different Ig superfamily proteins may thus act together to form an adhesive code for laminar specificity, but whether these proteins instruct laminar targeting or rather act to promote connectivity between matching synaptic partners within specific laminae is not yet clear. Homophilic interactions between matching pre- and postsynaptic partners alone seem insufficient to specify in which lamina such interactions occur, especially in the crowded environment of the IPL. Many of these Ig proteins can also interact with heterophilic ligands, which may contribute to encoding laminar specificity. In addition to heterophilic type II cadherin and homophilic Ig superfamily interactions, repulsive interactions also contribute to laminar specificity in the IPL (Box 1), indicating that the concerted action of many cues controls precise laminar connectivity in this system.

Subcellular specificity. The subcellular precise targeting of synapses is a key aspect of functional connectivity, as the location of a synapse on a target cell can have a profound influence on the output of that cell. The identity of the cell surface molecules that regulate subcellular synaptic specificity is still largely unknown, although a few examples involving type II cadherins and Ig superfamily proteins have now been described.

In the hippocampus, the proximal dendrites of CA3 pyramidal neurons in the stratum lucidum have complex spines, which are contacted by elaborate presynaptic boutons that originate from DG granule cells. The more-distal segments of CA3 dendrites have regular spines and receive inputs that do not originate from the DG. Both CA3 and granule cells express CDH9. In dissociated hippocampal cultures, knockdown of Cdh9 in CA3 neurons specifically decreases the number of DG-CA3 synapses, without affecting non-DG synapses onto CA3 neurons115. In vivo, knockdown of Cdh9 in granule cells reduces the number and size of DG synapses in the stratum lucidum, but laminar targeting of DG axons does not seem to be affected. Cdh9 knockdown in CA3 cells disrupts spine formation and results in the formation of long filopodia on the proximal region of CA3 dendrites. Because ectopic expression of CDH9 in cultured CA1 hippocampal neurons, which normally lack Cdh9, is not sufficient to increase the number of DG synapses onto those cells115, these findings are most-consistent with a permissive role for a homophilic CDH9 interaction in regulating the proper development of DG–CA3 synapses. In contrast to disruption of CDH9, disruption of CDH8, which is also expressed in granule cells and CA3 neurons, through treatment of organotypic slices with a peptide that interferes with CDH8 function does result in DG axon mistargeting116, suggesting that CDH8 and CDH9 might regulate different aspects of DG–CA3 connectivity.

The L1 family of Ig proteins has been implicated in subcellular-specific targeting in several circuits. A classic example is neurofascin, which contributes to the specific subcellular targeting of GABAergic basket cell axons to the axon initial segment of Purkinje cells117. A recent study provided evidence for a role of L1 family proteins and CNTNs in mediating subcellular specificity in the spinal cord, where a class of GABAergic interneurons called 'GABApre' neurons selectively target excitatory proprioceptive sensory terminals on motor neurons (Fig. 4). The remarkably precise subcellular targeting of GABApre axo-axonic synapses is mediated by a receptor complex of NRCAM and CHL1 on GABApre axons interacting heterophilically with a complex of CNTN5 and CNTN-associated protein 4 (CASPR4; also known as CNTNAP4) on sensory terminals118 (Fig. 2B). Loss of these Ig superfamily proteins, alone or in combination, results in a similar reduction in the density of GABApre boutons on sensory terminals (Fig. 4) but not in mistargeting of GABApre axons. These findings suggest that an Ig superfamily adhesive code contributes to determining the density of GABApre boutons on sensory terminals, and they imply that cells maintain subcellular compartment-specific distributions of surface proteins to guide the specificity of connectivity. Little is known about how the surface protein composition of subcellular compartments is organized and maintained. An example of a surface receptor with a subcellular-specific distribution is the LRR protein NGL2, which localizes to proximal hippocampal CA1 pyramidal cell dendrites via a trans-synaptic interaction with netrin-G2 expressed on Schaffer collateral axons innervating this dendritic segment119. Loss of Ngl2 reduces the density and plasticity of excitatory inputs on the proximal segment of the CA1 dendrite without affecting Schaffer collateral targeting or excitatory synapse density on the distal dendrite120,121.

Figure 4: An Ig superfamily combinatorial code regulates subcellular specificity in the spinal cord.
figure 4

a | Proprioceptive sensory neurons in the dorsal root ganglion form excitatory synapses on motor neurons in the spinal cord. These sensory nerve terminals represent the sole synaptic target of 'GABApre' interneurons located in the spinal cord. b | The remarkably precise subcellular targeting of these axo-axonic synapses is mediated in part by a complex of neuronal cell adhesion molecule (NRCAM) and close homologue of L1 (CHL1) on GABApre axons, and a receptor complex of the immunoglobulin (Ig) superfamily proteins contactin 5 (CNTN5) and CNTN-associated protein 4 (CASPR4) on sensory neuron terminals. c | Loss of these Ig superfamily proteins, alone or in combination, results in a similar reduction in the density of GABApre boutons on sensory terminals.

PowerPoint slide

Surface molecules and synaptic diversity

In addition to a highly specific wiring pattern, the properties of synaptic transmission critically contribute to information processing in the CNS. Presynaptic release probabilities, postsynaptic receptor compositions and plasticity vary widely across CNS synapse types. Molecular diversity could also contribute to generating unique synaptic properties, but little is known about the mechanisms by which surface proteins regulate synaptic diversity. Below, we highlight several examples of surface proteins regulating synaptic properties in the vertebrate CNS.

Neurexins regulate postsynaptic receptor composition. Alternative splicing of neurexins regulates their interactions with multiple synaptic ligands, which may contribute to the specification of synaptic properties. To directly test the in vivo significance of neurexin alternative splicing at SS4, a major regulator of neurexin–ligand interactions, one study genetically engineered mice in which SS4 was constitutively included in Nrxn3a and Nrxn3b transcripts, and could be excised by expression of Cre recombinase122. Constitutive inclusion of SS4 in Nrxn3 (leading to expression of NRXN3-SS4+) does not affect synapse number but selectively alters postsynaptic trafficking of the AMPA glutamate receptor (AMPAR), without affecting NMDA receptor trafficking. In neurons expressing NRXN3-SS4+, synaptic levels of AMPARs are decreased and AMPAR endocytosis is increased (Fig. 5a), resulting in a decrease in synaptic strength and impaired long-term potentiation (LTP). These effects are non-cell-autonomous and can be reversed by Cre recombinase-mediated excision of SS4, indicating that presynaptic SS4-lacking neurexins regulate postsynaptic AMPAR stabilization. The impaired AMPAR trafficking in neurons expressing NRXN3-SS4+ is associated with decreased surface expression of LRRTM2 (Ref. 122), which only binds to SS4-lacking neurexins45,46 and also interacts with AMPARs44. This study indicates that presynaptic neurexins trans-synaptically control postsynaptic AMPAR stabilization through an SS4-dependent interaction with LRRTM2 and possibly other ligands, such as neuroligins. The picture that emerges suggests that alternative splicing of neurexins modulates a cell surface interaction network with various ligands, thereby affecting AMPAR trafficking and changing postsynaptic receptor composition and the functional properties of synapses.

Figure 5: Neurexins and LRR proteins regulate the properties of synaptic transmission.
figure 5

a | Presynaptic neurexin 3 (NRXN3) trans-synaptically controls postsynaptic AMPA glutamate receptor (AMPAR) stabilization through a splice site 4 (SS4)-dependent interaction with leucine-rich repeat transmembrane neuronal protein 2 (LRRTM2) and possibly other ligands, such as neuroligin 1 (NLGN1). LRRTM2 only binds to SS4-lacking NRXN3 (NRXN3-SS4; only the α-isoform is shown) and also interacts with AMPARs (not shown). Constitutive inclusion of SS4 in NRXN3 (NRXN3-SS4+; only the α-isoform is shown) disrupts the interaction with LRRTM2 and selectively alters the postsynaptic trafficking of AMPARs without affecting NMDA receptor trafficking. In neurons expressing NRXN3-SS4+, synaptic levels of AMPARs are decreased and AMPAR endocytosis is increased, resulting in a decrease in synaptic strength and impaired long-term potentiation. b | The postsynaptic LRR protein extracellular LRR and fibronectin type III domain-containing 1 (ELFN1) regulates presynaptic release probability. CA1 axons form facilitating synapses with a low release probability on oriens lacunosum-moleculare (OLM) cells, and depressing synapses with a high release probability on parvalbumin-positive (PV+) cells. ELFN1 is expressed in OLM cells but not in PV+ cells. Knockdown of Elfn1 (not shown) in OLM cells decreases facilitation of CA1–OLM synapses, whereas ectopic expression of ELFN1 in PV+ cells moderately increases facilitation of CA1–PV+ synapses. These findings indicate that postsynaptic ELFN1 decreases presynaptic release probability, possibly via a trans-synaptic interaction with metabotropic glutamate receptor 7 (mGluR7; as indicated by the question mark).

PowerPoint slide

LRR proteins control presynaptic release properties. The above example illustrates how cell surface interactions between pre- and postsynaptic neurons regulate postsynaptic properties. Presynaptic release properties are also well known to depend on interactions between pre- and postsynaptic cells123,124. A striking example is found in the hippocampus, where CA1 pyramidal cell axons contact two types of interneurons: somatostatin- positive interneurons in the stratum oriens (oriens lacunosum-moleculare (OLM) cells) and parvalbumin- positive interneurons (PV+ cells) (Fig. 5b). The synapses onto OLM cells are facilitating (with a low release probability), whereas the synapses of the same CA1 axon onto PV+ cells are depressing (with a high release probability), indicating that the identity of the postsynaptic target cell defines these presynaptic properties (Fig. 5b). The postsynaptic LRR protein extracellular LRR and FN3 domain-containing 1 (ELFN1), which is expressed in OLM cells but not in PV+ cells, has a key role in regulating the presynaptic release probability of CA1 axons. Knockdown of Elfn1 in OLM cells decreases the facilitation of CA1–OLM synapses, whereas ectopic expression of ELFN1 in PV+ cells moderately increases the facilitation of CA1–PV+ synapses114, indicating that postsynaptic ELFN1 decreases the presynaptic release probability. The molecular mechanism by which ELFN1 regulates release probability is not completely understood. ELFN1 trans-synaptically recruits metabotropic glutamate receptor 7 (mGluR7)125, which displays a striking presynaptic enrichment at pyramidal cell–OLM synapses but not at synapses of pyramidal cells with other pyramidal cells or interneurons126. A high concentration of presynaptic mGluR7 could act to reduce neurotransmitter release at facilitating CA1–OLM synapses. Together, these studies illustrate how cell type-specific expression of a postsynaptic surface molecule instructs presynaptic molecular composition and functional properties.

Conclusions and perspectives

Molecular diversity of surface molecules has long been envisioned to mediate the specificity of connectivity in neural circuits. In recent years, substantial progress has been made in identifying cell type-specific expression patterns of molecularly diverse families of surface molecules and characterizing the complexity of their extracellular interactions. Despite this progress, we are still a long way from understanding how, and to what extent, molecular diversity contributes to the specific wiring of vertebrate neural circuits. Recent advances in single-cell profiling are beginning to support the existence of cell type-specific surface protein repertoires. For at least some cell types, in vivo evidence indicates that protocadherin diversity defines unique cell surface identities that are important for self-recognition. Other surface proteins contribute to this process as well, preventing the inadvertent formation of contacts with a neuron's own processes and thus indirectly contributing to wiring specificity. Whether unique repertoires of cell surface molecules also encode laminar, cellular and subcellular specificity, as well as synaptic diversity of individual neuron types is only beginning to be explored. The in vivo evidence for a role in specifying connectivity is probably strongest for the type II cadherins CDH8 and CDH9, whose cell type-specific expression instructs the laminar choice of BC axons in the retina. Even here, type II cadherins interact heterophilically with unknown ligands localized in the lamina, rather than in a 'lock-and-key' fashion with target cells, as envisioned by Sperry. The main role of type II cadherins in assembling this circuit may be to bias the choice of lamina, with additional surface molecules probably mediating the correct partner choice within that lamina. Much remains to be learned about the identity of the cues that govern lamina-specific innervation and how pre-patterned laminar environments develop127,128. Although cadherins, Ig and LRR superfamily proteins have now been shown to contribute to cellular and subcellular specificity in a few instances, we still know very little about the identity of the other cues regulating wiring specificity at these levels. Neurexin diversity does not seem to encode cellular specificity, although a role in subcellular specificity cannot yet be ruled out. The identity of the molecular cues instructing the final step of circuit assembly, synaptic diversity, is still largely unknown, although a few examples involving neurexins and LRR proteins have now been described.

The emerging picture is that no single cue regulates specificity, and that the same cue can be reused in multiple steps of circuit assembly. This suggests that many different surface molecules act in various combinations at each of the steps that lead to establishing precise connectivity. Combinatorial coding enhances robustness, increases information-coding capacity and, at the same time, limits the need for a vast number of different cues to encode specific connectivity127. In addition to requiring cell type-specific repertoires of surface molecules, such a combinatorial coding model implies that distinct subcellular compartments and synaptic subtypes will have different combinations of surface molecules. Unravelling how different combinations of surface molecules collaborate to build a functional nervous system constitutes a major future challenge that will need to be addressed in intact circuits and at specific synapse types. Determining how neurons dynamically alter their surface repertoires in response to activity, and how activity modulates protein interaction networks at synapses in vivo, will be another challenge. Bridging the gap between systems-level analyses of cell type-specific expression profiles or extracellular interactomes and the functional testing of candidate genes in connectivity will also be essential to arrive at a more-complete understanding of the molecular specification of synaptic connectivity.

Finally, many of the genes encoding cell surface molecules described in this Review have been linked to neurodevelopmental and psychiatric disorders30,58,129,130,131. Their cell surface localization makes these proteins valuable targets for interference with small molecules or drugs. A deeper understanding of the molecular mechanisms determining precise synaptic connectivity in neural circuits will therefore be essential to develop and realize future treatments for cognitive disorders.