Molecular and Cellular Mechanisms of Lamina-specific Axon Targeting

Lamination and Sublamination in the Vertebrate Retina

The inner plexiform layer (IPL) of the vertebrate retina illustrates basic features of layered brain architecture. The IPL is sandwiched between two cell body layers, the inner nuclear layer (INL) and the ganglion-cell layer (GCL). In the retina, the synapses between bipolar cells, amacrine cells, and ganglion cells are found exclusively in the IPL. Their density is enormous: every cubic micrometer of the IPL contains ∼100 synapses. Another set of retinal synapses, namely those formed between photoreceptors, horizontal cells, and bipolar cells are restricted to another very thin neuropil layer, the outer plexiform layer (OPL). This already precise targeting is made more complex by the fact that within each of these broad classes of cell types (amacrine, bipolar, and ganglion cells), there exist many subtypes. Indeed, there are between 50 and 120 morphologically distinguishable cell types in the vertebrate retina. The exact number depends on the species, with teleosts apparently holding the record (Badea and Nathans 2004; Connaughton et al. 2004; MacNeil et al. 1999; Mangrum et al. 2002; Masland 2001; Rockhill et al. 2002; Wagner and Wagner 1988). Each type has a characteristic (sub)laminar connection pattern in the OPL and/or IPL. Synapses between the 25–70 amacrine cell types, the 12 bipolar cell types, and the 15–20 ganglion cell types are stacked up within the IPL, forming at least eight anatomically separable sublaminae. These sublaminae are extremely thin: In the larval fish retina, they are just a few microns thick, barely accommodating a few synaptic terminals (Nevin et al. 2008). Frequently, but not always, dendrites and/or axons of retinal neurons are restricted to a single sublamina, giving rise to a single arbor with planar geometry. In fact, the vertical position(s) of branches within the neuropil layer, together with other morphological criteria, are frequently used to decide if two neurons belong to the same or different types. In the retinas of all vertebrates, there are many more cell types than anatomically distinguishable IPL sublaminae, arguing that a particular sublamina must necessarily be targeted by more than one cell type. Thus, there are at least three steps required to achieve synaptic specificity in the IPL: First neurites must be guided from the cell body into the neuropil; second, sublaminar targeting must occur; and, finally, selection of the synaptic partner amongst several cell types within each sublamina must take place. It is important to note that whereas bipolar cells and amacrine cells make layer-specific targeting choices strictly within the retina, ganglion cells make layer-specific choices for their dendrites in the retina, but also direct their axons out of the eye and into the visual brain areas, where they target particular neuropil layers, again in a type-specific fashion.

Layer-specific Photoreceptor Axon Targeting in Drosophila

The Drosophila visual system comprises the retina, and four optic ganglia, designated the lamina, the medulla, the lobula, and the lobula plate (reviewed in Meinertzhagen and Hanson 1993). As a compound eye, the retina is divided into an array of unit eyes, called ommatidia, each of which contains eight photoreceptors, R cells. Unlike vertebrate visual systems, in Drosophila, photoreceptor axons project directly into the brain, and display three distinct targeting preferences (Fig. 2A). Within each ommatidium, six R cells, designated R1–R6, extend axons into the outermost ganglion, the lamina, where they innervate a precise pattern of target cartridges such that each cartridge collects visual information from a single point in space, and neighboring cartridges are arranged to recreate a topographically correct representation of visual space (Braitenberg 1967; Kirschfeld 1967). These cartridges contain a reiterated array of projection neurons, designated L1-5 cells, which, in turn, project in an equally precise column and layer-specific fashion into the second optic ganglion, the medulla. Similarly, the axons of R7 and R8, the two remaining cells in each ommatidium, also innervate distinct layers within the medulla, within the same columns containing inputs from L cells that respond to visual signals from the same point in space. This targeting specificity emerges through a two-step process in which both R8 and R7 target precisely to a pair of temporary layers, before re-targeting to their final retinorecipient layers (Ting et al. 2005). Indeed, layer-specific targeting of many different neuron types is a broad organizational hallmark of the medulla, the lobula, and the lobula plate, and has been linked to specific behavioral functions in some cases (Bausenwein and Fischbach 1992; Lee et al. 2001). Moreover, the retinotopic arrangement of columns that is established initially by photoreceptor axons innervating the lamina and medulla is preserved across all optic ganglia. Overall, this system is characterized by a remarkable degree of precision, with axons making precise projections throughout development, leading to the establishment of an essentially invariant adult connectivity pattern.

Laminar Organization of the Vertebrate Retinotectal Projection

Retinal ganglion cell axons, the outputs of the vertebrate retina, make both retinotopic and layer-specific targeting decisions within a variety of brain regions. Of these regions, perhaps the most intensively studied is the optic tectum (as it is known in fish and birds, also known as superior colliculus in mammals). Here, ganglion cell inputs innervate a specific subset of morphologically and molecularly distinct layers (Inoue and Sanes 1997; May 2006; Nevin et al. 2008). Thus, the well-studied retinotectal map is, in fact, a stack of several superimposed maps. Laminar segregation of retinal axons is found in all vertebrates, although the exact order of layers (and particularly the nomenclature) strikingly varies between birds, fish, and mammals (Fig. 2 B–D). It is clear that retinorecipient tectal layers are specialized to receive their inputs from specific subtypes of retinal ganglion cells (Huberman et al. 2008b; Huberman et al. 2009; Xiao et al. 2005; Yamagata and Sanes 1995), although the lamina-specific projection patterns of many additional ganglion cell types remain to be characterized.

In mice, at least some of these layer-specific targeting decisions emerge gradually during development, with a specific subtype of retinal ganglion cells, the tOFF-aRGCs, initially projecting broadly across the depth of the SC before retracting to a specific lamina. Within their specific layer of the colliculus, these laminated projections are further organized into a regular array of columns that likely reflect the precise array of ganglion cell bodies in the retina (Huberman et al. 2008b).

In zebrafish, laminar specificity of RGC projections is precise from the earliest developmental stages. RGCs project their axons into six target layers, called the stratum opticum (SO) at the surface of the tectum, three sublayers of the stratum fibrosum et griseum superficiale (SFGS), the stratum griseum centrale (SGC), and a deep layer bordering the stratum album centrale (SAC) and the stratum periventriculare (SPV) (Xiao et al. 2005). Each RGC sends out a single axon, which may form collaterals outside the tectum, but, as a rule, innervates just one tectal layer (Xiao and Baier 2007). The Pou4f3 (Brn3c)-expressing subset (which is composed of several types) projects into the SO and into the two deeper sublayers of the SFGS. A Pou4f3:mGFP-expressing transgenic reporter line (see below) has been used extensively to study laminar targeting of axons in the tectum, including in forward genetic screens (Xiao et al. 2005), and of RGC dendrites in the retina (Mumm et al. 2006).

Formation of the Zebrafish IPL

Time-lapse imaging of fluorescently labeled amacrine cells and RGCs in the zebrafish retina has revealed the sequence of events that result in the formation of sublaminae (Fig. 3A–C). Here, amacrine cells are the first to extend neurites into the nascent IPL (Godinho et al. 2005). In fact, their processes form a band of neuropil in between their cell bodies, “pushing” one amacrine population into the future INL and another into the future GCL. The GCL-resident amacrine population will develop into the displaced amacrines of the mature retina. Sublaminar organization emerges almost instantaneously once neurites have joined up (Godinho et al. 2005). Although RGCs are the first-born neurons in the retina and begin to differentiate earlier than amacrine cells (Kay et al. 2001; Kay et al. 2005; Masai et al. 2000), their dendrites first accumulate in the space below the future displaced amacrines and then move past them to contact the amacrine neurites. RGCs are not required for amacrine sublamination, although their absence destabilizes and slows down this process, as seen in RGC-deficient lakritz/atoh7 mutants (Kay et al. 2004).

Mumm et al. (2006) reported that most RGC dendrites directly target the pre-existing amacrine sublaminae (Fig. 3B,C). The four early-developing RGC dendritic strata are formed roughly in an inside-out fashion. First, the proximal-most sublamina appears, then the adjacent one, just distal to the first, etc. The fifth stratum then breaks the rule and forms between the second and the third sublamina. Importantly, individual RGCs do not adhere to this stereotyped sequence; some project to more distal sublaminae immediately. Also, laminar allegiances of RGC dendrites can change over time, with some branches being retracted from one layer and re-established in a different layer. In fact, the lamination strategies of single RGCs are highly diverse, and it is unclear if this reflects true cell-type diversity or some variability in their developmental trajectories (Mumm et al. 2006). Finally, bipolar axons arrive in the IPL after amacrine and RGC processes have formed connections (Schroeter et al. 2006).

These detailed observations led to the prediction of several molecular and cellular mechanisms. First, amacrine neurites are attracted to each other and carry type-specific cell-surface cues that ensure homotypic fasciculation and perhaps heterotypic repulsion, giving rise to sublaminae by cell–cell recognition. This interaction does not require the presence of their main postsynaptic partners, the RGCs. Second, RGC dendrites are attracted to the amacrine plexus or could be independently polarized to extend their dendrites preferentially into an apical direction toward the amacrine plexus. Third, RGC dendrites read surface cues provided by the preformed amacrine scaffold to correctly target sublaminae. Fourth, the ongoing reorganization of RGC dendrites indicates that the cell–cell recognition systems must be dynamically regulated. Finally, because this dynamic dendrite behavior extends well into mature stages (after 10 dpf) when fish larvae already use their visual sense to capture prey and avoid obstacles, these events must occur in the context of a functional visual system.

Specific Cell-surface Adhesion Molecules

Perhaps the most intuitive model for how axons might choose their postsynaptic targets invokes the matched expression of homophilic or heterophilic cell-surface molecules on axons and dendrites that would allow axons to identify the correct postsynaptic partner from among a number of alternatives. This match could be at the level of a “unique” code in which only a single, homophilic molecule, or pair of heterophilic molecules, is matched between pre- and postsynaptic cells, or could be defined through a combinatorial mechanism in which different combinations of cell-surface molecules need to be assessed in order to identify the appropriate match. This notion was first put forward in the context of the diverse layer-specific patterns of expression of cell-surface molecules observed in the chick tectum (Miskevich et al. 1998; Wohrn et al. 1998). Experimental support has predominantly come from studies of neurite patterning in the IPL, but fragmentary evidence in favor of this hypothesis has also emerged from studies of the fly visual system and the vertebrate tectum.

In the last few years, four closely related cell-surface adhesion molecules, all members of the immunoglobulin-domain-containing superfamily, have been identified as being important for sublaminar target recognition in the chick IPL (Yamagata and Sanes 2008; Yamagata et al. 2002). These molecules, called Sidekick (Sdk1), Sdk2, Dscam, and DscamL, are expressed in nonoverlapping subsets of bipolar cells, amacrine cells, and RGCs, and are each shared by cells that project to the same sublamina from either side of the IPL. All four molecules bind homophilically, as shown in vitro, and are concentrated at the sites of synaptic contacts in the IPL. Overexpressing Sdk1 (or Dscam) in cells that are normally negative for these factors results in misrouting of its processes to the Sdk1- (or Dscam-) rich sublamina. Conversely, removal of these molecules by gene knockdown perturbs the lamina-specific wiring. These results strongly suggest that these IgSF molecules (and related factors waiting to be discovered) are important elements of a cell-surface recognition code that ensures precise connectivity in the chick retina.

A recent study tested whether Dscam and Dscam-like 1 regulate synaptic and laminar specificity in mammals (Fuerst et al. 2009). The authors used anatomical methods to evaluate laminar-specific targeting of several different amacrine and RGC subtypes as well as more direct measurements of synaptic specificity such as electron microscopy and paired electrophysiological recordings. All of these data support the conclusion that in the absence of one or both Dscams, retinal neuron recognize and form synapses with the appropriate partners. Especially remarkable is the fact that laminar and synaptic specificity still develops even though removal of Dscams causes retinal neurons to form dense aggregates with fasciculated dendrites (Fuerst et al. 2008, 2009). Thus, in mammals, Dscams regulate cell spacing but do not appear to mediate laminar-specific targeting or synaptic specificity as they do in the chick (Yamagata and Sanes 2008). The fact that appropriate synaptic partners still find one another in Dscam knockout animals suggests that adhesive cues do in fact regulate synaptic specificity in mammals. Still, the identity of adhesive cues that mediate overall target or cell-type-specific choices are unknown. Ango et al. (2004) established that an Ankryin-related Ig molecule Neurofascin mediates sub-cellular targeting of interneuron synapses in the cortex—one of the few examples of an adhesion-based mechanism for synaptic specificity in mammals. Despite the elegance and simplicity of an adhesion-based code for synaptic specificity, repellants undoubtedly play an essential role as well. A recent study demonstrated that semaphorin initiated repulsion dictates synaptic specificity of a subset of spinal neurons by restricting their axons away from certain postsynaptic targets (Pecho-Vrieseling et al. 2009). Such results are especially noteworthy because, in addition to providing a primary example of a molecular determinant of synaptic specificity in the mammalian CNS, they suggest that laminar specificity may emerge through signals that drive axons away from inappropriate targets.

As it is relatively difficult to manipulate molecules in the chick, efforts in recent years have focused on developing the mouse as a model system to examine the mechanisms underlying specific axon targeting. Indeed, different subfamilies of IgSF molecules have been shown to be important for the pathfinding to olfactory glomeruli (Kaneko-Goto et al. 2008; Serizawa et al. 2006) and even the targeting of particular subcellular domains on Purkinje neuron dendrites in the cerebellum (Ango et al. 2004). In the visual system, evidence for the involvement of any of these IgSF molecules is scarce. More progress has been made in the colliculus than in the retina. In particular, the identification of markers for a number of subsets of retinal ganglion cells (Huberman et al. 2008b; Huberman et al. 2009; Kim et al. 2008; Siegert et al. 2009; Yonehara et al. 2008) has opened the possibility of determining the extent to which each of these ganglion cell subsets displays layer-specific targeting, and the mechanisms by which these targeting choices are made. Indeed, using one of these RGC cell types, the tOFF-aRGCs, as a model, it was shown that lamina-specific targeting of these axons can be achieved even in the absence of retinal activity waves (which play critical roles in organizing other aspects of targeting) (Huberman et al. 2008b). As these neurons appear to first innervate a broad region of the colliculus before becoming restricted to their retinorecipient layer, molecular cues must regulate pruning of the inappropriate initial connections of these cells.

In the fly visual system, the layer-specific targeting choices made by R7 and R8 photoreceptors have provided the most direct insight into the molecular underpinnings of a “code-based” targeting decision. In particular, the leucine-rich repeat protein Capricious appears to play a critical, specific role in matching R8 axons to their appropriate target layer. Capricious is expressed in both R8 axons (but not R7 axons), and their target layer, and mediates homophilic interactions in vitro (Shinza-Kameda et al. 2006). In capricious mutants, R8 axons mistarget, extending into the R7 layer. Strikingly, mis-expression of capricious in R7 axons is sufficient to redirect their targeting specifically to the layer of the medulla normally innervated by R8. Finally, the expression of capricious by R8 but not by R7 appears to be an integral part of the cell fate determination of these two cells—a transcription factor called Senseless that controls the expression of many R8-specific genes regulates the capricious promoter directly, and is in turn specifically repressed in R7 cells (Morey et al. 2008). Thus, the matched expression of capricious between R8 axons and their prospective targets appears to play an important, instructive role in the targeting of these cells.

Lamination in the Drosophila Visual System by Dynamic Regulation of Cell Adhesion

In the fruit fly visual system, the molecular mechanisms underlying layer-specific targeting have been most closely examined in the context of the targeting choices made by R7, R8, and the L cells within the medulla. Strikingly, a number of lines of evidence show that a precisely timed pattern of expression of a relatively ubiquitous pair of cell-surface molecules, the classical cadherin, N-cadherin, and the receptor tyrosine phosphatase LAR appears to underlie the layer-specific targeting of a number of cell types. Both of these molecules were initially identified through studies of the developing Drosophila embryo, where embryos homozygous for mutations in either gene displayed a number of distinct defects in axon fasciculation, tract formation, guidance, and synapse formation (Iwai et al. 1997; Krueger et al. 1996). Given these broad, severe phenotypes, single cell somatic mosaic analysis of both genes in the visual system revealed the surprising result that R7 photoreceptors mutant for either of these genes displayed very specific defects in layer-specific targeting (Clandinin et al. 2001; Lee et al. 2003; Maurel-Zaffran et al. 2001). Further studies revealed that although mutations in these two genes have similar adult phenotypes, they play roles in distinct stages of the targeting process (Ting et al. 2005). Similar studies of N-cadherin and LAR in L cells revealed that N-cadherin plays both cell-autonomous and nonautonomous roles in directing the layer-specific branching patterns of specific cells (Nern et al. 2008). These latter studies also shown that changes in the fine-scale spatial and temporal pattern of N-cadherin expression in specific neurites correlated well with the phenotypes seen in mutant cells (Nern et al. 2008).

Direct evidence in favor of the importance of the timing of N-cadherin and LAR expression came through genetic studies of a transcription factor, Sequoia, which is expressed sequentially in R8 and R7 cells, and controls the timing of expression of both molecules (Petrovic and Hummel 2008) (Fig. 4B). In particular, sequoia mutant R7 and R8 cells target to inappropriate layers, while inducing prolonged expression of sequoia in R8 redirects these axons to the layer of the medulla normally innervated by R7. Strikingly, this effect of prolonging sequoia expression in R8 can be suppressed by removing N-cadherin, arguing that N-cadherin is a key effector. Together, these data argue that Sequoia controls the temporal pattern of N-cadherin and LAR expression, and that this regulation is at least part of a mechanism that controls the competence of R7 and R8 axons to target to specific layers of the medulla. More broadly, these studies show that relatively ubiquitous cell-surface molecules with broad patterns of expression in the nervous system can be used to direct specific targeting decisions through precise timing of their expression on different cell surfaces.

Unfortunately, there is, as yet, no direct evidence that precisely timed patterns of expression of cell-adhesion molecules plays an instructive role in layer-specific targeting decisions made by axons in vertebrate visual systems. Nonetheless, it is intriguing to note that N-cadherin plays a role in the development of layer-specific targeting of RGCs in the chick optic tectum. In particular, using a coculture system in which RGCs have simultaneous access to multiple tectal layers, a function-blocking antibody directed against N-cadherin was sufficient to alter layer-specific targeting choices (Inoue and Sanes 1997). Moreover, mutations in zebrafish N-cadherin affects lamination in the retina (Masai et al. 2003). Thus, dynamic regulation of such ubiquitous cell-adhesion molecules might be a broad theme of layer-specific targeting in a variety of experimental contexts.

Laminar Patterning by Basement Membrane-anchored Cues

The notion that the basement membrane on the surface of the embryonic tectum could serve as a “laminar organizer” has received experimental support from studies of the zebrafish dragnet mutation, which causes a subset of RGC axons to stray outside of their normal layer (Xiao et al. 2005). The mutation causes disruptions in the structure of the basement membrane lining the tectal surface and was mapped to col4a5, a gene encoding the network-forming collagen IV α5 (Xiao and Baier 2007). Intriguingly, RGC axons in which Integrin β1, an essential component of the main collagen IV receptor, was knocked down, showed normal targeting, suggesting that the axons are not responding directly to collagen IV. Rather, the guidance activity is likely due to one or more secreted factors that are bound to the basement membrane. Importantly, the effects of this mutation are phenocopied by enzymatic and genetic manipulations of heparan sulfate proteoglycans, molecules that have been shown to act as cofactors for a variety of guidance molecules in other contexts. Thus, the basement membrane provides, directly, or indirectly, guidance cues to developing RGCs that are necessary to achieve appropriately restricted layer-specific targeting.

Laminar Patterning by Extracellular Matrix Factors

Studies in the optic tectum have revealed a number of cues, deposited in the extracellular matrix, that might help to organize layer-specific axonal targeting or other aspects of synaptic specificity. Among them are the large glycoproteins Tenascin-C, Tenascin-R, and Reelin, as well as the chondroitin sulfate proteoglycan Versican (Yamagata and Sanes 2005). Bands of Tenascin-R immunoreactivity border the retinorecipient layers, and Tenascin-C and Tenascin-R have been shown in vitro to be repellent factors to RGC axons. These results suggest that the function of Tenascins may be to prevent axons from crossing into inappropriate territory (Becker et al. 2004; Perez and Halfter 1993). Loss of Reelin function in mouse and rat results in misrouting of axons that is restricted to the superficial retinorecipient layers of the tectum, where Reelin is expressed (Baba et al. 2007; Sakakibara et al. 2003). It is unclear if the observed defect is specific to laminar targeting, as this phenotype is obscured by disruptions of tissue architecture, which are caused by earlier functions of this secreted molecule in cell migration. However, a function for Reelin in circuit formation, as opposed to cell positioning, has been established in the retina, where AII amacrine cells express Dab1, an essential component of the Reelin signal transduction machinery. In both reeler and dab1 mutants, aberrant connections are formed in the IPL (Rice et al. 2001).

Most recently, the concept of extracellular laminar positional cues has been boosted by the identification of Nel, a glycoprotein containing five chordin-like and six epidermal growth factor-like domains (Jiang et al. 2009). In the developing chick tectum, expression of Nel is localized in specific laminae that retinal axons normally do not enter. Although the receptor is unknown, Nel-binding activity is detected on retinal axons both in vivo and in vitro. In vitro, Nel inhibits retinal axon outgrowth and induces growth cone collapse and axon retraction. These results suggest a role for Nel in the establishment of laminar specificity in the retinotectal projection.