Developmental regulation of axon branching in the vertebrate nervous system

Our ability to perceive, to act and to remember is a reflection of the elaborated synaptic connections of neural circuits. The development of these connections relies on the proper regulation of axon branching, a morphogenetic process that allows a single neuron to communicate with multiple partners via its only axon. Since the pioneering observations by Ramon y Cajal (Ramon y Cajal, 1904), axonal branches have been found throughout the vertebrate nervous system. They help to define the morphology and connectivity of each neuronal cell type and provide alternative strategies for target selection and structural plasticity (O'Leary et al., 1990; Yamahachi et al., 2009). Therefore, understanding the mechanisms that underlie the temporal and spatial control of branching is crucial in the study of neural circuit development.

Axonal branches appear in different shapes, form at different locations, and change with time. Some are long and some are short; some are large, forming exuberant arbors that cover the entire target, whereas some are small, innervating just a small region of the target; some are dense and some are sparse, with only one branch projecting from the main axonal shaft; some are made during early development and form major neuronal pathways that persist throughout life, whereas others form at late stages of development and can be remodeled by neural activity. How are these branches generated? How is such diversity achieved? How is their development tailored to their function? How are they shaped by evolutionary pressures in modern animals? These are fascinating, yet challenging, questions in developmental neurobiology.

To address these questions, it is practical to consider several common branching processes that are often mentioned in the literature. They can be characterized by the morphology, complexity and function of the branches they generate (Fig. 1). One form of branching is arborization (see Fig. 1A and Glossary, Box 1), which typically occurs at axon terminals in the target region and results in the formation of tree-like arbors with higher-order branches. As discussed in more detail below, examples of arborization are found in retinal ganglion cells and sensory neurons. By contrast, branches with the simplest shape are made by bifurcation (Fig. 1B), a process that often generates two daughter branches that grow away from each other, as seen in the central sensory projections in the spinal cord. Between these two extremes are collateral branches (Fig. 1C), which usually form far from the nerve terminals (see Glossary, Box 1). They extend either orthogonally or obliquely from the axon, and often project to targets that are different from that of the main axon. This form of branching is exemplified by sensory collaterals in the spinal cord and by the descending projections from the cortex (O'Leary and Terashima, 1988).

Studies of these branching processes have recently been carried out in various vertebrates, including fish, amphibians, birds and rodents. With the aid of technological advances in neuronal culture,

molecular genetics and cell imaging, some common principles of their developmental regulation have begun to emerge. In this review, we focus on several well-studied branches in the peripheral (PNS) and central (CNS) nervous systems and illustrate how branching morphologies are shaped by extracellular factors (see Table 1) and neural activity. We discuss the unique features of axon branching and explain why it involves more than just forming a branch and why it is more complex than axon growth and guidance. We also highlight the critical steps that are shared among different branching processes during embryonic and postnatal development. We hope that, in combination with concomitant advances in our knowledge of the intracellular regulation of branching (for reviews, see Dent et al., 2003; Kornack and Giger, 2005; Schmidt and Rathjen, 2010), this review will bring us closer to understanding an important developmental process in the assembly of complex neural circuits in vertebrates.

Axonal branches in mature neural circuits result from the tightly regulated and stepwise cellular processes that occur during development, but their final morphology is not always indicative of how they form. The best illustration of this is the formation of axonal arbors of retinal ganglion cells (RGCs; see Glossary, Box 1) at the terminal zone of the superior colliculus (SC), a brain region in chick and rodents that receives visual inputs from the retina (Fig. 2). In early development, RGC axons initially overshoot their respective terminal zones (see Glossary, Box 1). Later, interstitial branches (see Glossary, Box 1) form preferentially in the region of the appropriate terminal zone and then become terminal arbors after their overextended axons retract (Fig. 2B) (Nakamura and O'Leary, 1989; Simon and O'Leary, 1992). Thus, the RGC terminal arbors are not necessarily initiated at axon terminals but instead arise from previously formed interstitial branches, suggesting that the mechanism that underlies the development of axonal branches can be very different from that implied by their final morphology.

Recent genetic studies (Ma and Tessier-Lavigne, 2007; Schmidt et al., 2009; Schmidt et al., 2007; Zhao and Ma, 2009; Zhao et al., 2009) of the bifurcation of dorsal root ganglion (DRG) sensory axons in the dorsal root entry zone (DREZ) of the spinal cord (see Fig. 3 and Glossary, Box 1) further support this view. These studies suggest that the generation of such a simple structure involves multiple steps that are regulated by different extracellular cues (Fig. 4). The first step is directly linked to the formation of the daughter branch and requires C-type natriuretic peptide (CNP; also known as NPPC), a hormone that binds to the membrane-associated guanylate cyclase natriuretic peptide receptor 2 (NPR2) and stimulates the production of cyclic guanosine monophosphate (cGMP). The genetic ablation of the CNP precursor gene Nppc, its receptor Npr2 or the downstream cGMP target cGMP-dependent protein kinase G1 (Prkg1), each leads to a failure in the formation of the second branch (Fig. 4B). This defect suggests that branch formation is regulated by the CNP pathway during bifurcation. This conclusion is supported by studies of dissociated DRG neurons in culture, in which CNP treatment, pharmacological activation of cGMP and overexpression of Prkg1 all promote branch formation (Zhao and Ma, 2009; Zhao et al., 2009).

The second step involves the guidance of the two daughter branches by the guidance molecules SLIT1 and SLIT2, acting on their receptors ROBO1 and ROBO2 (Ma and Tessier-Lavigne, 2007). In mutant mice lacking both ligands or both receptors, the DRG central afferents (see Glossary, Box 1) still bifurcate but no longer have the normal orientation of the T- or Y-shaped bifurcation fork, as in about half of the neurons one daughter branch follows the normal trajectory whereas the other enters the spinal cord (Fig. 4B). This phenotype suggests that SLITs are normally required to keep the sensory afferents outside of the spinal cord and guide them to the DREZ, presumably through repulsion. This repulsive function for SLITs is supported by the relatively gentle collapsing activity found for SLIT2 on sensory growth cones in culture (Ma and Tessier-Lavigne, 2007).

The above phenotypes suggest several interesting features of bifurcation. First, the two daughter branches are not created equally at the beginning, as one daughter branch develops as an extension of the primary afferent, whereas the other branch forms under the regulation of CNP. Second, these two branches have different molecular and biochemical properties as their growth direction is regulated differently. One of them is influenced by SLITs, whereas the other is not. In addition, there seems to be no preference for which branch becomes the ascending or the descending projection, as either one can be affected in the above-mentioned mutants.

Based on these features, we propose a simple repulsion-coupled collateral formation model for sensory axon bifurcation (Fig. 4C). After the sensory afferents reach the DREZ they encounter several environmental cues and respond with the following steps, which might occur very rapidly. First, a stop signal from SLITs, and possibly from other repulsive factors, blocks the growth of the afferents at the DREZ but does not turn them away from it. Instead, the same signal steers the growth cones into the longitudinal track, either anteriorly or posteriorly in a random fashion. Then, CNP stimulates the formation of a new branch at the turning point and promotes the growth of this new branch. Finally, both daughter branches grow in opposite directions along the DREZ, forming the ascending and descending projections. This model suggests that the formation of the second branch might be similar to the formation of collateral branches, rather than involving the growth cone splitting mechanism associated with the final morphology, as previously thought (Acebes and Ferrus, 2000).

Although further studies are needed to validate this model, an intriguing extension of the above discussion is that any branching process, including those commons forms, can be subdivided into several distinct developmental steps. These steps, including branch formation, growth, guidance and pruning, might constitute the core mechanism of axon branching; coordination of their regulation might be sufficient to generate the diverse branching pattern by specifying the location, number, angle/trajectory, size and complexity of branching. For example, collateral formation is dictated by the location of branch formation, followed by guidance of the newly formed branch, whereas for axon bifurcation branch formation is tightly coupled to the guidance of both the nascent branch and the main axon. In addition, arborization could simply be the result of the repetition of these processes in association with branch interactions, such as self avoidance and tiling (see Glossary, Box 1). Therefore, understanding these developmental steps is crucial to elucidating the mechanisms of axon branching. In the following sections we discuss how these steps are regulated by extracellular cues and by neural activity to generate stereotyped branches during development. The cellular mechanisms of their regulation (see Box 2) have been reviewed recently (Dent et al., 2003; Kornack and Giger, 2005; Schmidt and Rathjen, 2010) and will therefore only be discussed in the appropriate context.

Branches can form at nerve terminals or interstitially along axonal shafts, depending on where branches are initiated. Several mechanisms that specify the location of branching have emerged from recent studies of the extracellular factors involved in branching (Fig. 5).

Most axonal branches form at nerve terminals in their target regions, as exemplified by both the peripheral and central axons of sensory neurons in the DRG (see Fig. 3). The terminal arbors of these neurons develop after the axons have reached their targets, such as the peripheral skin tissue or the central motoneurons in the spinal cord. Several target-derived factors have been found to provide instructive signals that specify where an axon will arborize.

One such signal is nerve growth factor (NGF), the first neurotrophic factor (see Glossary, Box 1) implicated to be acting as a target-derived signal from the skin to promote terminal branching of sensory axons in the peripheral tissue (Kennedy and Tessier-Lavigne, 1995). To separate its role in cell survival from axon branching, a double deletion of Ngf and Bax was created to block programmed cell death in mice (Patel et al., 2000). In this mutant, the peripheral projections reach their skin targets but fail to innervate them and to arborize. Although this phenotype might reflect the function of NGF in axon growth, studies of dissociated Bax^–/– neurons in culture, in which they survive in the absence of neurotrophic factors but do not develop a branched morphology, indicate that branch elongation and arborization depend on NGF and other neurotrophic factors such as neurotrophin 3 (NT3; also known as NTF3) (Lentz et al., 1999). This notion is further supported by reduced sensory arborization in the peripheral tissue in mice that lack the low-affinity NGF receptor p75 (also known as NGFR) (Bentley and Lee, 2000).

A role for NGF in axon branching has also been demonstrated in neurons from the sympathetic ganglion (see Glossary, Box 1). Its deletion in a Bax null mouse background results in the reduced innervation of a variety of sympathetic targets owing to deficient axonal branching (Glebova and Ginty, 2004). Many of these phenotypes are target dependent, and not all sympathetic targets have the same degree of reduced innervation, indicating the presence of other axon branching factors. A recent study suggests that WNT5A might be a good candidate to mediate NGF-induced branching (Bodmer et al., 2009).

Another Wnt family protein, WNT3, has been implicated in regulating central terminal branching of proprioceptive sensory neurons in the mouse spinal cord (Krylova et al., 2002). These terminals synapse with motoneurons in the ventral spinal cord where Wnt3 is expressed (Krylova et al., 2002). Growing dissociated embryonic mouse DRG neurons in the presence of WNT3-conditioned medium or next to ventral mouse spinal cord explants increases the number of secondary and higher-order branches, revealing the potential role of WNT3 as a target-derived signal that specifies the location of terminal arborization.

In addition, factors in the target tissue can also regulate the location of bifurcation (Fig. 5A). As discussed above, CNP is expressed in the mouse dorsal spinal cord at the time when sensory axons initially reach the DREZ (Schmidt et al., 2009; Zhao and Ma, 2009), and thus provides a local cue in this intermediate target to specify the location of bifurcation of sensory axons (Fig. 4). It is interesting to note that only two branches form here instead of the exuberant arbors found at the mature DRG axon terminals, indicating that CNP might act differently to other target-derived cues or that other factors are present to restrict the number of branches formed.

Factors from local tissues can provide an instructive cue to promote branch formation along the axon (Fig. 5B). This was first demonstrated by the study of axons that project from layer five of the rat cortex to their subcortical targets through the formation of collateral branches (O'Leary and Terashima, 1988). The target tissues appear to secrete an attractive diffusible factor that promotes branch formation (Heffner et al., 1990; Sato et al., 1994).

Another example of locally induced branching is the development of sensory collaterals that sprout from the ascending and descending projections (see Fig. 3B). During early spinal cord development these branches do not form immediately after sensory afferents bifurcate (Mirnics and Koerber, 1995b; Ozaki and Snider, 1997); instead, there is a 2-day delay before the collaterals sprout. In vitro studies suggest that factors are present in the spinal cord that might be upregulated to directly stimulate the formation of these collaterals (Wang et al., 1999).

Results from in vitro studies indicate that several extracellular molecules can act as local cues to promote branch formation along an axon. One example is NGF, which initiates collateral sprouting from axons of cultured chick DRG neurons when it is presented locally on coated beads (Gallo and Letourneau, 1998). A similar response has been observed to fibroblast growth factor 2 (FGF2) in embryonic hamster pyramidal neuron axons, but in this case the factor acts in close proximity to the growth cone (see Glossary, Box 1) (Szebenyi et al., 2001).

The guidance molecule netrin 1 can also stimulate branch formation of cultured hamster cortical axons when applied locally through a pipette tip (Dent et al., 2004). Filopodial protrusions were initiated at smooth segments of these axons in the vicinity of the pipette tip, followed by filopodial extension towards the tip. The effect of netrin 1 on branching is possibly mediated by calcium signaling, as netrin 1 treatment induces spatially restricted calcium transients in the axon that coincide spatiotemporally with new branch formation (Tang and Kalil, 2005). Moreover, the local application of netrin 1 induces a local Ca²⁺ transient that is accompanied by growth of the stimulated branch (Hutchins and Kalil, 2008).

Finally, the secreted protein anosmin (also known as KAL1) may promote the formation of local collateral branches from projection neurons in the mammalian olfactory bulb. Anosmin is defective in Kallman syndrome, a human disease that is associated with anosmia and linked to the absence of the olfactory tract. In culture, anosmin has been shown to stimulate branch formation in rat olfactory neurons (Soussi-Yanicostas et al., 2002).

Another mechanism that restricts branch formation to certain regions of the axon has been suggested from studies of retinotopic map development (see Glossary, Box 1) in the SC of chick and rodents (see Fig. 2), where interstitial branches form along the axon near the appropriate terminal zone (Luo and Flanagan, 2007; McLaughlin and O'Leary, 2005). A study of chick RGC axons has indicated that the same ephrin/Eph gradients used for map formation might also participate in the generation of topographic-specific arbors (Yates et al., 2001). In culture, the axons of chick RGCs derived from the temporal side of the retina show a distinct preference for forming branches on membrane stripes that are derived from the anterior tectum, their appropriate terminal zone, whereas nasal RGC axons show no preference. The addition of soluble EPHA3 receptors to the culture to sequester ephrin A ligands on the membrane stripes completely abolished this biased branch formation, suggesting that ephrin A-mediated inhibition can restrict branching locally. This notion is consistent with the in vivo observation of ectopic terminal zones in the posterior SC of mice that lack ephrin A2, ephrin A5, or both (Feldheim et al., 2000).

However, this ephrin A/EphA mechanism alone is not sufficient to generate posterior-specific branching by nasal RGC axons. A recent investigation in mice suggests that a complementary gradient of EPHA7 exists in the target structure to regulate this phenomenon (Rashid et al., 2005). Epha7 is expressed in a decreasing gradient along the collicular anterior-posterior (A-P) axis, whereas ephrin A is more highly expressed on nasal RGC axons than on temporal axons, forming a decreasing gradient along the retinal nasal-temporal axis (Fig. 2A). Loss of Epha7 in mice leads to the formation of a densely branched ectopic terminal zone in the anterior portion of the SC, in addition to the normal posterior terminal zone (Rashid et al., 2005). Thus, in contrast to the ephrin A forward signaling (see Glossary, Box 1) that is employed in restricting temporal RGC axon branching, ephrin A reverse signaling (see Glossary, Box 1) serves as a negative regulator of nasal RGC axon branching. Thus, a common feature of RGC axon branching is that inhibition provides a key mechanism for generating topographic-specific branches.

Such a mechanism might also apply to the topographic axonal arborization of dentate gyrus granule cells in mouse hippocampal formation (Galimberti et al., 2010). The axons of dentate gyrus granule cells, termed `mossy fibers', project to the CA3 region of the hippocampus, where they form one or more terminal arborizations at locations in CA3, based on the position of the granule cell soma in the dentate gyrus. The dentate gyrus exhibits an expression gradient of the EPHA4 receptor, which suggests that it could participate in the topographic specification of mossy fiber terminal arborizations. Indeed, disrupting EPHA4 signaling in mouse hippocampal slice cultures increases the number of mossy fiber terminal arborizations, in addition to abolishing the topographic specificity (Galimberti et al., 2010).

However, branching inhibition only restricts where branches form, but does not guarantee eventual branch formation. It is possible that axons are intrinsically capable of making branches, as demonstrated by branch formation of chick RGCs in culture (Yates et al., 2001). In addition, studies of the Xenopus optic tectum found that a branch-promoting signal provided by brain-derived neurotrophic factor (BDNF) can complement ephrin A/EphA-mediated branch restriction (Cohen-Cory and Fraser, 1994; Cohen-Cory and Fraser, 1995). BDNF is uniformly expressed in the tectum, and its high-affinity receptor TrkB is uniformly expressed in the retina (Cohen-Cory and Fraser, 1994). Injection of BDNF into the tectum increases the branching and complexity of RGC terminal arbors, whereas injection of specific neutralizing antibodies to BDNF reduces axon arborization and complexity (Cohen-Cory and Fraser, 1995). Furthermore, a recent in vitro study of chick neurons suggests that this branch-promoting activity is mediated by an interaction between TrkB and ephrin A at the cell surface (Marler et al., 2008). TrkB interacts with ephrin A5 via a cysteine-rich domain, and overexpression of this domain in chick RGCs results in a dominant-negative effect on BDNF-induced branch formation. Similarly, addition of soluble EPHA7 to the RGC culture inhibits this activity, as do membrane stripes studded with EPHA7 (Marler et al., 2008). These experiments suggest that a global branch-promoting cue, such as BDNF, coordinates with the ephrin/Eph gradients to control RGC axon branching. Thus, the coupling of global branch promotion with a complex branch-restricting system provides a general molecular mechanism to restrict branching to a specific location in the developing nervous system (Fig. 5C).

A defining feature of axonal branches is the angle between the daughter branch and the parent axon. This can be determined by the intrinsic properties of each axon (Katz, 1985; Lasek, 1988), but might reflect the interaction of newly formed branches with the environment. Similar to axon guidance, extracellular factors can guide the growth trajectory of daughter branches through both short-range and long-range guidance mechanisms that influence the branching angle and the final branching pattern. As seen during sensory axon bifurcation discussed above, SLITs control the growth direction of the daughter branches (Fig. 4C). In the absence of SLIT signaling, the normal orientation of the bifurcation fork is altered, resulting in one daughter branch entering the spinal cord (Fig. 4B). However, only half of the afferents are affected in these mutants, suggesting the presence of other extracellular signals that guide branch growth during bifurcation (Ma and Tessier-Lavigne, 2007).

The invasion of sensory collaterals into different laminar targets in the spinal cord provides an illustration of how different populations of branches are regulated by the environment. For example, cutaneous afferents from small diameter neurons that convey pain and temperature information are stopped at the dorsal lamina in the rodent spinal cord, whereas muscle afferents from large diameter neurons course through the entire spinal cord and reach the motoneurons in the ventral half of the spinal column (see Fig. 3B) (Mirnics and Koerber, 1995b; Ozaki and Snider, 1997). Interestingly, the muscle afferents follow a hyperbolic trajectory and appear to be attracted directly to the motor column, while avoiding the dorsal gray matter (Snider et al., 1992).

Another example of environmental regulation is provided by RGC axon branches, which, in addition to preferentially forming near the appropriate terminal zone, have a strong preference to extend in either the lateral or medial direction, depending on their relative location (Fig. 2B) (Nakamura and O'Leary, 1989). This preference might be controlled by ephrin B/EphB signaling in mice, as the EPHB receptor is expressed in an increasing dorsal-ventral (D-V) gradient, whereas the ligand ephrin B1 is expressed in an increasing lateral-medial (L-M) gradient in the

SC (see Fig. 2A) (Hindges et al., 2002). Furthermore, after the normal period of refinement has passed, mutant mice that lack functional EPHB retain ectopic terminal zones at positions lateral to the correct terminal zones as a result of an increase in laterally oriented branches, coupled with a decrease in medially oriented branches (Hindges et al., 2002). A second study using retroviral transfection showed that graded ectopic expression of ephrin B1 results in the formation of laterally oriented branches within the ectopic terminal zone, possibly owing to branch repulsion by the high level of ephrin B1 present on the medial side of the ectopic zone (McLaughlin et al., 2003a).

Thus, the branching angle is determined by the guidance of newly formed branches by mechanisms that are similar to those of axon guidance. Factors present in the local environment or acting from target regions can attract or repel each branch and thereby influence branch morphology and function.

Axonal branches that form during development are not always retained in the mature neural circuit (Luo and O'Leary, 2005). Many can be eliminated through an active process called pruning. The classical example of pruning is the layer-five projection neurons from the motor and visual cortex (O'Leary and Koester, 1993). During early mammalian cortical development, multiple collaterals form from these projections, but some are eliminated during postnatal development to make the appropriate connection with their targets. RGC development provides another example of pruning (Fig. 2B), as both the overshooting axons and the inappropriately formed interstitial branches are rapidly eliminated over a period of a few days (McLaughlin et al., 2003b).

Branch pruning may involve cellular factors that are important for branch degeneration (Luo and O'Leary, 2005), and its regulation may employ the same molecules that are used for axon guidance (Vanderhaeghen and Cheng, 2010). For example, an early study of the SEMA3A receptor, plexin A3 (PLXNA3), suggests that pruning of specific hippocampal mossy fibers and of pyramidal branches requires this pathway (Bagri et al., 2003), as both pruning events are defective in the Plxna3 mouse knockout. Moreover, a recent study has suggested that mossy fiber pruning can be mediated by ephrin B/EphB reverse signaling (Xu and Henkemeyer, 2009). In addition to such molecular regulation, an activity-dependent mechanism has been proposed for branch pruning in RGCs as well (see below).

Terminal arbors are responsible for the innervation of the target field with a high degree of fidelity. Their morphology influences the function and plasticity (see below) of neural circuits and is often defined by branch number, length and order, as well as by arbor size. These parameters are commonly used to distinguish different neuronal cell types and to describe changes in arbor morphology during development. They reflect the interplay of branch formation, growth and guidance, as well as interactions between branches.

Recent studies of the peripheral projections of sensory neurons from the DRG and of the trigeminal ganglion (TG; see Glossary, Box 1) have illustrated how arbor size is regulated by competing factors in the environment. One factor is the SLIT family of secreted proteins, which can positively regulate the peripheral branching of TG axons in mice (Ma and Tessier-Lavigne, 2007). In wild-type embryos, the TG axons form a major arbor above the eye, but in Slit2;Slit3 double mutants the arbor is greatly reduced in size owing to a reduction in both branching and growth, whereas in Slit1;Slit2;Slit3 triple mutants the entire arbor is missing. The same defect was also found in embryos that lack the SLIT receptors ROBO1 and ROBO2. These results are consistent with a study in zebrafish embryos, in which slit2 overexpression dramatically increased the branching of peripheral sensory axons (Yeo et al., 2004).

This positive role of SLITs is counterbalanced by the semaphorin (SEMA) family of guidance molecules, particularly SEMA3A. These molecules are known for their activity in inhibiting axon growth and repelling growth cones (Dickson, 2002). Genetic studies have revealed enhanced peripheral axon growth and branching in both DRG and TG neurons from mouse embryos that lack SEMA3A (Kitsukawa et al., 1997; Taniguchi et al., 1997). Similar defects are found in mice in which the genes encoding the SEMA3A co-receptor neuropilin 1 or plexin A3/A4 have been deleted (Gu et al., 2003; Yaron et al., 2005). Such an activity is consistent with the inhibitory activity of SEMA3A on axon branching of dissociated cortical neurons in culture (Dent et al., 2004). Therefore, the negative regulation by SEMA3A cooperates with the positive regulation by SLITs to determine the arbor size of peripheral sensory axons.

The repulsive action of SEMAs and of other inhibitory cues might also play a role in self avoidance and tiling, two processes that reflect the interactions between branches within the terminal arbors. Live in vivo imaging of sensory axons in zebrafish has shown that removing one axon arbor allows the neighboring neurons to expand their arbor (Sagasti et al., 2005), indicating the existence of a repulsive mechanism to prevent the formation of overlapping receptive fields. Although the molecular mechanisms that underlie these processes have been extensively studied in the dendritic development of non-vertebrates, it would not be surprising if some of the factors identified there play similar roles in axonal tiling and avoidance in vertebrates (Grueber and Sagasti, 2010). For example, the cell adhesion molecule DSCAM that regulates these processes in axons and dendrites of Drosophila neurons (Grueber and Sagasti, 2010) has recently been implicated in dendritic avoidance in the rodent retina (Fuerst et al., 2009; Fuerst et al., 2008) and might well regulate axonal arbor morphology in vertebrates.

Developing neurons experience patterns of activity – both spontaneous and experience dependent – that contribute to the initial development and subsequent refinement of the functional circuit (Spitzer, 2006). Following the classic work of Wiesel and Huber (Wiesel and Hubel, 1963), many early studies convincingly demonstrated that activity is required for the segregation of visual inputs to produce so-called `eye stripes', or regions of the target structure that are innervated by axons carrying information from one eye and are adjacent to regions innervated by axons carrying information from the other eye (Antonini and Stryker, 1993; Reh and Constantine-Paton, 1985; Shatz and Stryker, 1988; Sretavan et al., 1988). Importantly, the loss of segregation in the absence of neural activity is the direct result of increased arborization by the afferent terminals, demonstrating that activity can potently influence branched structures. More recent work has revealed that activity-dependent axon branching is mediated by a variety of mechanisms, including the relative rates of branch addition and retraction, competitive interactions between neighboring branch arbors, and synapse-dependent branch stabilization (Fig. 6). These different mechanisms are likely to cooperate to produce the final arbor morphology and provide multiple means by which axon branching can be influenced by neural activity.

Axonal arbors often form as a result of a dynamic process that consists of the ongoing addition and retraction of nascent branches. These so-called `branching dynamics' can be influenced by neural activity to produce a bias toward branch addition or retraction, thus contributing to the final arbor size and shape (Fig. 6A).

Thalamocortical (TC) axons (see Glossary, Box 1) typically arborize early in mammalian development in cortical layer four to establish contacts with their target cells. An early study in cats showed that blocking neuronal activity with tetrodotoxin (TTX) significantly reduces the area occupied by TC axonal arbors, in addition to reducing the total branch number (Herrmann and Shatz, 1995). Recent studies have investigated this further using rat organotypic slice cultures in which high levels of spontaneous activity are present (Uesaka et al., 2007). TC axons in a thalamic explant grown next to the ventral surface of a cortical explant innervate and preferentially arborize in their normal layer-four target. Time-lapse analysis of fluorescent protein-labeled axons in culture indicates that arborization is a result of the bias in branching dynamics toward branch addition in the target layer. Furthermore, blocking all neuronal activity with pharmacological inhibitors causes arbor growth to decrease significantly, owing to a reduction in branch addition but not in branch extension.

An earlier study of rat cortical neurons that arborize in layer 2/3 of the cortex is also consistent with this conclusion (Uesaka et al., 2005). Axons traced by fluorescent proteins in cortical slices formed branches that are similar to those found during normal development in vivo, again as a result of branching dynamics being biased toward branch addition. However, the number of branches was reduced by pharmacological treatments that block neuronal activity, probably as a result of reduced branch addition (Uesaka et al., 2005). Thus, the biased regulation of branching dynamics can lead to a net increase or decrease in arbor size and complexity by shifting the dynamics toward branch addition or branch retraction, respectively. These slice culture results must be interpreted with care, however, as they are derived from ex vivo preparations in which many long-range inputs have been lost, and as such might not recapitulate development in vivo.

The branches of neighboring axons often innervate a shared or adjacent target field. Initially, there is an overabundance of branches, some of which are then retracted to produce mutually exclusive innervation fields. In addition to self avoidance and tiling, there is increasing evidence that this developmental pruning occurs via a competition-based mechanism in which `winning' axons are retained and `losing' axons are retracted (Fig. 6B).

Studies of zebrafish retinas in which subsets of RGCs are deficient in either electrical or neurosecretory activity have provided evidence that activity contributes to the interactions between neighboring RGC terminal arbors as they compete for limited target space (Gosse et al., 2008; Hua et al., 2005). Specifically, individual RGCs with suppressed activity via the expression of an exogenous potassium channel or of a dominant-negative SNARE protein show decreased axon terminal growth and reduced formation of new terminal arbors, but this inhibition can be relieved if the activity of neighboring RGCs is also suppressed (Hua et al., 2005). A more recent zebrafish study complements this finding using an elegant transplant approach to produce retinas that contain only a single RGC (Gosse et al., 2008). In the absence of any competition from neighboring RGCs, and thus no inhibition of branching, these single RGCs develop terminal arbors that occupy a greater area of the target and are more complex than RGCs in wild-type animals (Gosse et al., 2008). Thus, activity-dependent competitive interactions between neighboring RGCs may restrict arbor size and complexity.

This model is consistent with an earlier observation made in mouse RGCs that lack the β2 subunit of the nicotinic acetylcholine receptor (CHRNB1), which is required for the generation of spontaneous activity in developing retinal neurons prior to eye opening. These mutant RGCs fail to refine their terminal arbors, indicating that the absence of spontaneous activity eliminates any potential activity-dependent competition between neighboring RGCs, allowing their terminal arbors to grow unfettered by neighboring inhibition (Grubb et al., 2003; McLaughlin et al., 2003b).

However, a very recent study in zebrafish (Fredj et al., 2010) using tetanus toxin to suppress synaptic activity suggests that this activity-dependent branch regulation might be more complex. Here, the loss of activity in a single RGC led to enlarged terminal arbors that invaded the territory of neighboring RGCs, whereas suppressing the activity of neighboring RGCs at the same time resulted in normal terminal arbors. This effect is opposite to that found by Hua et al. (Hua et al., 2005), which might be due to the different approaches used to suppress synaptic activity. Nonetheless, activity-dependent competition is a critical determinant of RGC terminal arbor morphology.

A potential molecular mechanism that mediates activity-dependent competition has been suggested from studies of rodent sympathetic neurons (Singh and Miller, 2005; Singh et al., 2008). In culture, electrically stimulated sympathetic axon collaterals exhibit a distinct growth advantage compared with unstimulated collaterals. This effect appears to be mediated by BDNF, which is produced by electrically active axons in response to depolarization, as well as by its receptor p75, which is upregulated in unstimulated axons. Moreover, both BDNF and p75 are required for sympathetic axon competition in vivo. In wild-type mice, single sympathetic neurons of the superior cervical ganglion (SCG), the source of sympathetic innervation of the head, initially innervate two regions of the eye but are then pruned so that each neuron innervates only a single region. In p75^–/– mice this pruning does not occur, resulting in single neurons that innervate two compartments and single compartments innervated by multiple neurons (Singh et al., 2008). Similarly, mice bearing an activity-insensitive mutant Bdnf gene show the same deficit. Interestingly, the axon pruning appears to be the result of axon degeneration, as opposed to simple retraction, and this process requires BDNF and p75. Thus, these findings suggest a model (Fig. 6B) in which the most active axonal branches – the `winners' – secrete factors such as BDNF, which then act on the less active neighboring branches – the `losers' – to directly induce axonal degeneration and regulate branch stability (Singh et al., 2008).

Early work in the cat convincingly demonstrated that neural activity is required for the remodeling of terminal arbors (Antonini and Stryker, 1993). Occluding one of the eyes to block visually evoked neural activity resulted in a dramatic reduction in the complexity of geniculocortical terminal arbors (see Glossary, Box 1) over a period of only a few days. Interestingly, the time frame of this reduction correlates well with the period over which the physiological response to the deprived eye is lost in the cortical neurons, suggesting that synapses on the retracted branches are lost and that the remaining synapse-bearing branches can rapidly remodel. Thus, the regulation of synapses provides another mechanism by which neural activity may influence arbor morphology.

Such a mechanism has been examined in a number of studies of the retinotectal terminal arbors in Xenopus and zebrafish. The relationship between synaptogenesis and branch remodeling was first demonstrated by a study in Xenopus tadpoles, in which time-lapse imaging of fluorescent protein-labeled synaptobrevin II was used to visualize synapses in RGC axons (Alsina et al., 2001). This revealed that during the period of normal remodeling, most synapses are relatively stable; however, new synapses are both added and eliminated as the terminal arbor is reshaped. Interestingly, new branches usually originate at synaptic sites, whereas retracted branches tend to have few synapses present. This dynamic correlation may involve BDNF, which has been shown to modulate activity-dependent branching (Cohen-Cory, 1999). Microinjection of BDNF into the tectum increased both arborization and synapse number, but had stronger effect on synapses, suggesting that synapse formation might directly contribute to the regulation of axon branching (Alsina et al., 2001). In addition, neutralizing endogenous BDNF with function-blocking

antibodies eliminates both synapses and branches, indicating the involvement of BDNF in the stability of both structures (Hu et al., 2005).

Like BDNF, another extracellular factor, netrin 1, has also been implicated in regulating both synaptogenesis and branching in Xenopus (Manitt et al., 2009). Microinjection of netrin 1 protein into the tectum increased both branch dynamics and synapse formation, whereas blocking the netrin 1 receptor DCC with antibodies specifically prevented the addition of new branches and synapses, without affecting existing branches and synapses. However, the manner in which these two molecular cues – and any as yet unidentified cues – coordinate their actions to regulate RGC terminal arborization and synaptogenesis remains to be determined.

Two recent studies have further explored the intimate relationship between branch regulation and synaptogenesis by live imaging of zebrafish and Xenopus retinotectal projections (Meyer and Smith, 2006; Ruthazer et al., 2006). Using fluorescently labeled synaptophysin, a synaptic vesicle protein in the presynaptic compartment, both studies followed the maturation of developing synapses based on label intensity and simultaneously examined the behavior of developing terminal branches. New branches were found to preferentially form in the region of highest fluorescence intensity, presumably reflecting the presence of mature synapses, whereas new synapses were preferentially added to new branches. In addition, stabilized branches consistently had bright puncta along their length, indicative of the presence of mature synapses, whereas mature synapses halted retracting branches (Meyer and Smith, 2006; Ruthazer et al., 2006).

Furthermore, because RGC terminal arbors are continually remodeled on a small scale in response to visual experience, the Xenopus study also tested the structural response to evoked activity using patterned visual stimulation (Ruthazer et al., 2006). This paradigm increased the stability of the branches that bear intense puncta and induced the retraction of branches with faint or no puncta, strongly suggesting that experience-dependent activity is capable of regulating the retention of branches with mature synapses while inducing the retraction of branches with unstable synapses (Ruthazer et al., 2006). Interestingly, however, very recent work in zebrafish suggests that the presence of a synapse, but not its activity, is important for branch stabilization, as individual neurons that have been silenced by tetanus toxin are still capable of generating a normal number of stable branches (Fredj et al., 2010). Thus, the effect of visual stimulation is likely to be due to an effect of sensory-induced activity on the maturation of synapses, rather than on their function.

Taken together, these studies demonstrate that the maturity of developing synapses can regulate axon branching by promoting the growth of nascent branches and determining the selective stability of extended branches, and that visual activity can induce terminal arbor remodeling in a synapse-dependent fashion, providing a mechanism by which ongoing sensory experience can regulate axon branching (Fig. 6C).

The formation of proper axonal branches is a critical step in the establishment of functional neural circuits. The studies that we have discussed here illustrate important concepts in the developmental regulation of axon branching. Although some ideas need to be tested further (see Box 3), the complex regulation of axon branching can be broken down into several developmental steps, which include branch formation, growth, guidance and pruning, as well as branch interactions, such as competition, self avoidance and tiling. The spatiotemporal coordination of these key steps by both extracellular factors and neural activity at different locations on the axon could potentially generate the many types of branches found in the vertebrate nervous system.

Since these principles are derived from several well-studied neurons with stereotypic branched morphology, future studies will need to validate them in other types of neurons and branches. In addition, many extracellular cues have roles in the multiple processes discussed above. Distinguishing these activities requires a fuller understanding of the intracellular signaling mechanisms of axon branching, such as those involving GSK3 and PTEN, as well as their connections to the cytoskeleton (see Box 2) (Dent et al., 2003; Drinjakovic et al., 2010; Zhao et al., 2009). Furthermore, the intrinsic properties of neurons can influence axonal behaviors and modulate their responses to extracellular cues. Detailed cell biological studies of branch formation will provide a complete picture of how branches develop, how genetic programs dictate branching regulation, and how activity modulates branches. Finally, although different in structure and function, axons and dendrites both develop branches and might share molecular and cellular mechanisms of branching that are evolutionarily conserved (Grueber and Sagasti, 2010; Jan and Jan, 2010). Recent studies of axonal and dendritic development in flies and worms (Alexander et al., 2010; Hao et al., 2010; Jan and Jan, 2010; Oren-Suissa et al., 2010) will thus provide useful insights into the many outstanding questions regarding the regulation of axon branching in vertebrates (see Box 3).

Axon branching is crucial to vertebrate circuit development and contributes to guidance, targeting and plasticity (O'Leary et al., 1990; Yamahachi et al., 2009). The detailed molecular mechanisms identified here and in the future will provide useful tools to test circuit function. More importantly, many neurological and psychiatric disorders have been associated with defects in synaptic connections during development, which may result from the perturbation of branching regulation due to both genetic and environmental factors (Geschwind and Levitt, 2007). Indeed, several molecules described above have already been linked to these diseases, including ROBO in autism and anosmin in Kallmann syndrome (Anitha et al., 2008; Franco et al., 1991; Legouis et al., 1991). Understanding their regulation will provide new insights into the development and etiology of these disorders. Finally, branch sprouting has been suggested to be a viable means for regenerating nerve connections following injury (Cafferty et al., 2008), and understanding how they are formed during development will help to identify new ways to stimulate functional recovery in injured adults.

We thank Whit Tao and Avraham Yaron for reading the manuscript and the anonymous reviewers and the editor for constructive comments and suggestions. This work is supported by grants from the NIH and the Whitehall Foundation. Deposited in PMC for release after 12 months.