Networks are an important aspect of our daily lives. But I am not talking about social networks. I am talking about the networks that make us who we are and what we do: neural networks. Neural connectivity currently is a buzz word in neuroscience. Clearly, we need to understand the formation and function of neural networks in order to explain the pathology of neurodevelopmental and neuropsychiatric disorders. Disorders such as intellectual disability, autism spectrum disorders, or schizophrenia have their roots in aberrant neural circuit formation ^{1– 4} . But neural circuit formation is a multi-step process starting with cell differentiation and migration, involving axon guidance and synaptogenesis, and ending with synaptic maturation or pruning, a process that is not fundamentally different from the one ensuring synaptic plasticity in the adult nervous system. Therefore, we still struggle to understand the molecular mechanisms of neural circuit formation. In this review, I will concentrate on only one aspect of neural circuit formation: the navigation of axons to their target cells.

During the last 25 years, we have learned a lot about the molecular mechanisms underlying the process of neural circuit formation. Initial studies focused on muscle innervation ^{5– 7} and wiring of the visual system ^{8, 9} . Soon, dorsal commissural neurons in the spinal cord became the prime model for axon guidance studies because molecular mechanisms can best be studied in a simple system with clear readouts. Subsequently, the molecular mechanisms identified in these neurons have been tested in the establishment of more complex neural circuits in the brain ^{10– 16} .

Axons that have to form long-distance connections are cutting down their pathfinding into manageable steps; that is, they use intermediate targets on their way to the final target. These intermediate targets, or choice points, are crucial for axonal navigation by providing guidance cues. Guidance cues are molecules deposited in the extracellular matrix or expressed on cells along the pathway of navigating axons. They can be subdivided into short- and long-range guidance cues. Long-range guidance cues indicate the overall direction of growth but do not specify the actual pathway. This is done by short-range guidance cues. Both types of guidance cues can be subdivided into attractive and repulsive molecules.

For the dI1 subpopulation of commissural neurons in the spinal cord ^{17, 18} , we know axon guidance cues and their receptors for all four categories. At the lumbar level of the spinal cord, dI1 commissural axons have a very stereotypic trajectory. Axons extend ventrally toward the floor plate, the ventral midline of the spinal cord. They cross the midline and then turn rostrally along the contralateral floor-plate border. At least some of these axons target the cerebellum ¹⁹ . Long-range guidance cues guide them to their intermediate target, where short-range guidance cues take over to make sure that axons cross the midline before turning into the longitudinal axis. First, dI1 axons are guided ventrally by long-range repellents derived from the roof plate, the cells at the dorsal midline: BMP7 ^{20– 22} and Draxin ²³ . At the same time, axons are attracted toward the ventral midline, the floor plate, by Netrin1 binding to Dcc and Neogenin ^{24– 27} . In addition, axons are attracted toward the floor plate by Shh binding to Patched and Boc ^{28, 29} . Axons enter the floor plate, mediated by interactions between Contactin2 and NrCAM ^{30, 31} . Based on screens in flies, the repulsive midline-associated guidance cues, the Slits and their receptors (the Robos) were then identified as the reason why axons leave the intermediate target and why they never turn back ^{32– 38} . In addition to Slit, class 3 Semaphorins repel post-crossing commissural axons and prevent them from re-crossing ^{39– 42} . Finally, morphogens were identified as the guidance cues directing post-crossing commissural axons toward the brain ^{43– 48} .

So, case closed? Do we know everything about axonal navigation of the midline? Actually, far from it. Despite the identification of all of these molecules, we still do not fully understand how axons cross the midline and why they turn rostrally. This is not only because additional guidance cues for the navigation of the spinal cord midline were identified ^{40, 49, 50} but also because the regulation of the different receptors is not clear. Axons do not linger at the midline before they move on. Thus, the switch from attraction to repulsion has to be timed very precisely. Obviously, if axons were not attracted to the intermediate target, they would not get there; but if they did not start to be repelled upon arrival, they would not leave and move on. At the same time, they have to be equipped with receptors for the detection of guidance cues for the longitudinal axis. These receptors are not supposed to react to the guidance cues for the longitudinal axis on the ipsilateral side of the floor plate, despite the presence of the gradients of guidance molecules on both sides of the midline. So how is this precise timing of responsiveness achieved? Are similar mechanisms responsible for the formation of more complex circuits in the brain? Why so many cues for a seemingly rather simple decision at a choice point?

In the beginning of the molecular era of axon guidance, the focus was on the discovery of new families of axon guidance cues and their receptors ^{51– 53} . This has clearly changed. In recent years, the discovery of new guidance cues has been a rare event, but it has still happened. For instance, lyso-phosphatidyl-Β-D-glucoside, a glycerophospholipid, has been identified as a subpopulation-specific guidance cue for sensory afferents in the dorsal spinal cord ⁵⁴ . PRG2/Lppr3 (plasticity-regulated gene-2), a molecule interacting with lysophosphatic acid, was shown to regulate guidance of thalamocortical axons ⁵⁵ . But clearly, the focus in axon guidance research today is on the characterization of regulatory mechanisms. Axons are not guided by static interactions between guidance cues and their receptors on the growth cone. Rather, the responses of axons to their environment are induced by dynamic changes of these interactions. At the level of guidance cues, changes can lead to the synergistic activation of signaling pathways, as shown, for instance, for Netrin and Ephrin signaling during hindlimb innervation ⁵⁶ . Pre-crossing axons are attracted to the floor plate by a co-operation between Shh and Netrin ⁵⁷ .

The interaction between receptors can result in silencing or enhancement of a response. For example, in contrast to commissural axons, which are attracted by floor-plate-derived Netrin1, motoneurons are repelled by Netrin1 ⁵⁸ . Netrin1 is attractive for axons expressing Dcc but repulsive for axons expressing both Dcc and Unc5. Furthermore, the interaction between Dcc and Robo1 in commissural axons was shown to silence the attractive response of Netrin on post- compared with pre-crossing axons ⁵⁹ . More recently, many more of these receptor interactions in the plane of the growth cone membrane have been identified as important regulators of growth cone behavior (see “Some guidance receptors are regulated at the protein level” section).

Another problem that has become, and still is, a focus of axon guidance research is the aspect of timing. Precise timing of the switch from an attractive to a repulsive behavior is, of course, required at any choice point, but again it is best illustrated with the example of midline crossing. Clearly, the switch in axonal behavior is caused by a change in the expression of guidance receptors on the growth cone surface. This is a seemingly simple change, but the abundance of possibilities that is actually found is surprising ( Figure 1). Theoretically, expression of guidance receptors on the growth cone surface can be modified by changes in gene transcription, translation, or protein transport as well as protein stability. In fact, all of these possibilities have been confirmed experimentally to occur in axons crossing the midline either in the spinal cord or in the visual system.

Endo- and exo- cytosis of guidance receptors have been demonstrated to be required for a growth cone’s response to repulsive or attractive guidance cues, respectively ^{80– 83} . However, beyond these mechanistic aspects, specific changes in trafficking of guidance receptors have been shown to be involved in the temporal control of receptor expression ⁸⁴ . First demonstrated in flies, the expression of Robo on the growth cone surface was demonstrated to be dependent on specific trafficking in a Commissureless-dependent ^{85– 87} but Rab-independent ⁸⁸ manner. In vertebrates, a Commissureless ortholog is not found, but trafficking as a means to control surface expression of Robo is nonetheless maintained in vertebrates ^{38, 89} . In fact, the possibility to transport guidance receptors to the growth cone surface by shuttling them wrapped in subpopulations of vesicles not only provides specificity but also allows for an efficient change of the repertoire of guidance receptors, as large numbers of molecules can be inserted upon a stimulus that triggers the fusion of a specific subset of vesicles. A recent study has demonstrated the presence of Robo1 and Frizzled3 guidance receptors in different subpopulations of vesicles that were trafficked in a Calsyntenin1-dependent manner ⁸⁹ . Calsyntenin1 acts as a linker between the vesicles with specific cargo and the kinesin motor. RabGDI appears to be part of the stimulus that regulates the timing of vesicle trafficking, but only for those vesicles carrying Robo1 as cargo. The vesicles containing Frizzled3, the receptor for Wnt signaling regulating antero-posterior guidance of post-crossing commissural axons, also depend on Calsyntenin1 for transport, but RabGDI is not required for the delivery of Frizzled3 to the growth cone surface.

Although we may know most axon guidance molecules today, we still need to learn much more about axon guidance mechanisms before we can explain how neural circuits form. This is even true for simple circuits, such as those formed by spinal cord interneurons! In particular, we will need to understand how the interactions of guidance receptors with their ligands trigger specific intracellular signals and how they are translated into changes in growth cone behavior. While we start to understand individual signaling pathways, we do not have a clear idea how they interfere with each other. Furthermore, temporal changes in signaling are poorly understood. We have identified roles for individual molecules during different developmental processes contributing to the formation of neural circuits ^{1– 3} , but it is still unclear how binding partners or the downstream signaling (or both) change over time. So we will need not only to test molecules and mechanisms identified in simple circuits in complex brain circuits but also to conceptually and experimentally integrate different signaling pathways to understand the behavior of growth cones and axons during neural circuit formation. The analysis of single-gene knockouts is not sufficient to understand the dynamic role of a protein during neural circuit formation. Eventually, we will have to study neural circuit formation at the protein level. This remains a challenge because tools are not yet available to visualize protein function in vivo. We can follow molecules in isolated cells in vitro with sophisticated high-resolution imaging methods. We can also image neuronal activity in vivo in actively behaving animals. But there is a huge gap in between that needs to be closed before we understand formation and function of neural circuits. However, the development of new tools and the refinement of existing technology will allow us to tackle some of these open questions in a variety of model systems and thus to assemble more and more pieces of the puzzle.