From whole organism to ultrastructure: progress in axonal imaging for decoding circuit development

How our brain develops and operates has captivated the mind of scientists for centuries. Our knowledge in neuroscience was notably revolutionized at the dawn of the 20th century when Ramón y Cajal settled a long-standing debate regarding the fundamental architecture of the nervous system. Using a staining method developed by his ideological opponent, Camillo Golgi, Ramón y Cajal showed neurons to be discrete cells rather than a continuous, interconnected network (Ramón y Cajal, 1991; Sotelo, 2002). This theory, the neuron doctrine, became a fundamental principle of neuroscience and prompted the advent of studies aimed at understanding how neurons form complex, yet precise, neural circuits. During development, neurons extend highly dynamic axons that interpret guidance cues in their environment to navigate towards their final synaptic targets. This pathfinding is ensured by motile structures at axonal ends, the growth cones, that respond to the numerous extracellular signals they encounter via receptors present on their surface. Axons often extend great distances from the soma and adopt distinct behaviors to ensure the accuracy of circuit wiring. Axons not only grow, pause and turn, but also fasciculate (bundle) with each other, repel one another and retract or degenerate after taking erroneous paths (Poulain and Chien, 2013; Spead and Poulain, 2020; Tessier-Lavigne and Goodman, 1996; Wang and Marquardt, 2013). Early light microscopy and electron microscopy (EM) studies have highlighted the morphological features of individual neurons and their axonal projections in various tissues. In contrast, recent advances in imaging technologies, including light-sheet fluorescent microscopy (LSFM) and super-resolution microscopy (SRM), have provided insights into axonal development in unparalleled detail, enabling the reconstruction of neural circuits in vivo and the discovery of the cellular and molecular mechanisms governing circuit development (Maglione and Sigrist, 2013). In this Review, we highlight major advances in the field of axon development driven by technological progress in microscopy. We first present major contributions and limitations of EM and conventional light microscopy to our understanding of developing axons. We then introduce recent findings in axon growth, guidance and circuit wiring achieved via LSFM and SRM. We highlight how these new technologies have provided an improved, integrated view of circuits and developmental processes in vivo. We further describe breakthroughs in defining molecular signaling pathways that govern axon growth and guidance, focusing on signal integration at the axonal membrane and cytoskeletal dynamics. We finally explore how emerging technologies can push the boundaries of axonal imaging to further decode the intricate mechanisms of circuit development in vivo.

Since Ramón y Cajal's seminal work on brain architecture, developmental neurobiologists have taken advantage of constantly improving labeling and imaging techniques to dissect the morphology of neurons and the cellular organization of complex circuits across species. Axonal projections were first visualized by light microscopy in fixed tissues after filling neurons with a dye, allowing the reconstruction of neuronal tracts and patterns of connectivity from serial sections. This technique was notably used in the 1960s and 1970s to trace axons in various ganglia of small invertebrate models (Davis, 1970; Kerkut et al., 1970; Stretton and Kravitz, 1968). Serial electron micrographs also largely contributed to these early mapping efforts, leading to the reconstruction of the entire nervous system in C. elegans and Drosophila (White et al., 1986; Zheng et al., 2018), and the first characterization of fine intra-axonal components such as neurofilaments and microtubules (Wuerker and Kirkpatrick, 1972). Later on, the introduction of improved anterograde and retrograde tracers combined with fluorescent immunolabeling broadened the characterization of anatomical circuits by enabling the simultaneous visualization of multiple axonal tracts together with identified neurons (Gerfen and Sawchenko, 1985; van der Kooy and Steinbusch, 1980). Coupled with the development of high-resolution confocal laser scanning microscopy in the 1990s, these tracing and labeling approaches were instrumental for delineating axonal projections in vertebrates at different stages of development and their organization into topographic maps in sensory systems (Harris, 1986; McLaughlin and O'Leary, 2005). Axons were also found to adapt to the environment they navigate in, correcting their trajectories after embryological manipulations (Guthrie and Lumsden, 1992). Concomitant progress in genetics with the advent of gene targeting and genetic screens enabled the identification of signaling factors required for proper axon guidance and circuit formation. Ephrins, for example, were discovered as the main guidance cues governing the targeting of retinal axons in the optic tectum (Frisen et al., 1998; Nakamoto et al., 1996).

Neuro-anatomical tracing and EM drove our current understanding of axonal development and circuit wiring but nonetheless present shortcomings, the most notable being their incompatibility with live imaging. In addition, some axonal structures do not survive the fixation process during EM and are therefore difficult to identify from a micrograph (Leterrier et al., 2017; Letourneau, 2009). The discovery of fluorescent proteins together with the development of molecular and viral tools to label specific cells, structures or proteins addressed these challenges and opened a new era in the study of developing axons. It became possible not only to visualize axon and growth cone dynamics within an organism (Higashijima et al., 2000; Knobel et al., 1999; Murray et al., 1998; Zallen et al., 1999), but also to monitor and track intra-axonal processes sustaining axonal behavior. Anterograde transport of cytoskeletal proteins, for example, was known to exist from pulse-chase radiolabeling experiments conducted in the 1970s (Hoffman and Lasek, 1975; Willard et al., 1979) but only became fully appreciated after neurofilaments tagged with green fluorescent protein (GFP) were observed to have rapid anterograde movements separated by long pauses in live cultured neurons (Roy et al., 2000; Wang et al., 2000). This ‘Stop and Go’ model of slow axonal transport was also found to apply to microtubules (Wang and Brown, 2002). GFP and cytoskeletal regulators fused to fluorescent proteins were also expressed in developing neurons in C. elegans and visualized by in vivo time-lapse confocal imaging, leading to the identification of essential signaling pathways regulating growth cone organization and protrusive activity during axon formation and guidance (Adler et al., 2006; Chang et al., 2006; Quinn et al., 2008; Norris et al., 2009). Finally, the use of fluorescent proteins enabled the visualization and functional characterization of receptors at the surface of growth cones and axons (O'Donnell et al., 2009). For example, fusions of the receptor Robo to GFP or its pH-sensitive variant pHLuorin have been used to monitor the targeting and trafficking of Robo at the surface of navigating axons in vivo, leading to the discovery of molecular mechanisms fine-tuning the responsiveness of axons to the guidance cue Slit (Keleman et al., 2005; Pignata et al., 2019). Altogether, decades of research in developmental neurobiology has exploited the major advances in tracing, genetic engineering and imaging mentioned above to paint the detailed picture of axonal development and circuit wiring we have today. We have characterized the outgrowth and trajectories of many axonal tracts during development in various species, identified most guidance factors instructing how circuits are built, and deciphered many intra-cellular cascades underlying growth cone responses to guidance cues (Dudanova and Klein, 2013; Gorla and Bashaw, 2020; McCormick and Gupton, 2020; Spead and Poulain, 2020; Stoeckli, 2018; Suárez et al., 2014). Yet many questions remain unanswered due to limitations in imaging techniques. Although confocal microscopy has clear advantages for live-imaging axons, it is restrained by a diffraction limit of ∼200 nm (Maglione and Sigrist, 2013), which prevents the resolution of individual axons in a densely packed nerve bundle, let alone axonal structures separated by only a few tens of nanometers. We are thus still limited in our understanding of how individual axons elongate, communicate with one another or change their behavior to form an entire nervous system. Similarly, although we have identified cues, receptors and intra-axonal processes regulating axon guidance, we do not know how these factors and structures operate and are spatially and temporally regulated at the molecular scale. Progress in these areas require imaging modalities capable of nanometer-scale resolution, large fields of view and live analysis that we are only beginning to achieve and have only recently led to major breakthroughs.

Deciphering how individual axons develop into precise neuronal networks in vivo requires improved spatial and temporal resolution to observe fine structural details and fast dynamic processes, all the while limiting photobleaching and phototoxicity for live imaging. Recent advances in genetic, imaging and computational processing have addressed these challenges and opened a new era in the study of circuit wiring by enabling the visualization of single axons in vivo and the reconstruction of entire connectomes.

A major advance in our understanding of circuit wiring came with the development of genetic approaches that specifically label multiple individual axons within a brain structure, thereby providing a unique way to map neuronal connectivity over long distances with remarkable resolution. Among them, Brainbow and its derivatives proved to be especially powerful as they use a recombinase-mediated DNA inversion or excision strategy to stochastically express a unique combination of green, red and blue fluorescent proteins within a cell (Cai et al., 2013; Lichtman et al., 2008; Livet et al., 2007; Weissman and Pan, 2015). As the combination of these three primary colors generates all colors in the visual spectrum, each Brainbow-expressing neuron displays a unique, distinct color that allows visualization and tracing of its axonal projection. Combined with confocal microscopy, the Brainbow approach was first used to visualize individual oculomotor and hippocampal mossy fiber axons in the adult brain, and to reconstruct 341 mossy fiber axons and 93 of their postsynaptic targets, the granule cells, in the cerebellum (Livet et al., 2007). Multicolor tracing with Brainbow has since been adapted to multiple organisms and used in various systems to delineate the precise architecture of circuits and identify underlying wiring mechanisms (Egawa et al., 2013; Hadjieconomou et al., 2011; Hampel et al., 2011; Heap et al., 2013; Pan et al., 2011; Robles et al., 2013). The recent reconstruction of 71 retinal ganglion cell (RGC) projections in the dorsal lateral geniculate nucleus, for example, led to the discovery that multiple retinal axons converge onto relay cells in that region, challenging the textbook notion that each relay cell receives input from only one or two RGCs in the adult (Hammer et al., 2015).

Concomitant with the development of Brainbow and other labeling approaches, progress in microscopy has tremendously increased both the acquisition speed and three-dimensional resolution of imaging, thereby allowing the reconstruction of axonal projections in exquisite detail (Table 1). LSFM notably differs from confocal microscopy in that it illuminates a sample with a thin sheet of light perpendicular to the axis of fluorescence detection (Voie et al., 1993). The coincidence of the light sheet with the focal plane of the fluorescence detection system restricts the excitation of fluorophores to the illuminated plane, thereby avoiding phototoxicity or photobleaching and enabling fast imaging of large biological specimens from multiple different directions over long periods of time (Huisken et al., 2004; Keller and Stelzer, 2008; Tomer et al., 2012). Axial resolution and sensitivity of LSFM are determined by the thickness of the light sheet, which varies depending on how the sheet is generated. For example, sub-diffraction limit resolution has been achieved using ultrathin optical lattices to image both non-neuronal cells (Chen et al., 2014; Gao et al., 2012; Wan et al., 2019) and neurons (Corsetti et al., 2019; Lemon and Keller, 2015; Liu et al., 2018). Different methods can be used to create light sheets, and we refer interested readers to excellent recent reviews for more technical details (Albert-Smet et al., 2019; Hillman et al., 2019; Wan et al., 2019). Thanks to its advantages, LSFM has rapidly been employed in combination with solvent-clearing methods for imaging neural circuits in whole embryos with unprecedented resolution (Mano et al., 2018; Ueda et al., 2020). Opaque embryos rendered transparent with solvent-based clearing procedures such as 3DISCO [three-dimensional imaging of solvent cleared organs (Ertürk et al., 2012)] or its more-recent derivatives iDISCO or uDISCO (Pan et al., 2016; Renier et al., 2014) can be immunostained and imaged, allowing the precise tracing of axonal projections together with their molecular identification. When used on mouse transgenic embryos, this approach has led to the three-dimensional reconstruction of motor and sensory projections with unparalleled precision, enabling the visualization of fine motor axon terminal arborizations from individual forelimb nerves (Belle et al., 2014; Liau et al., 2018). More recently, 36 cleared human embryos and fetuses ranging from 6 to 14 weeks of gestation were immunostained with over 70 antibodies and imaged with LSFM to generate the most advanced three-dimensional atlas of human development at the cellular resolution (Belle et al., 2017). The entire human peripheral nervous system was reconstructed over time, providing a precise developmental time-lapse of the sensory innervation of the human hand (Fig. 1). Interestingly, a striking variability in the innervation of the left and right hands was observed, with the left and right ulnar and radial nerves harboring noticeable differences in branch number and length (Belle et al., 2017). Heterogeneity in wiring was similarly observed in Drosophila using LSFM combined with expansion microscopy, a method whereby a biological sample is physically enlarged using an isotopically expanding polymer (Gao et al., 2019). Long-range tracing of olfactory projection neurons elongating from the antenna lobe to the calyx of the mushroom body and lateral horn revealed subcellular wiring details that varied among samples analyzed. The size and number of boutons formed in the calyx notably differed among neurons from a same hemisphere as well as between animals (Gao et al., 2019). Thus, by enabling the visualization and three-dimensional reconstruction of axonal trajectories in vivo at high resolution, LSFM has brought to light previously uncharacterized anatomical details of circuits and their variability across individuals.

Although reconstructing precise connectomes from cleared samples provides invaluable insight into the developmental trajectories that axons take to form circuits (Friedmann et al., 2020), it does not provide a dynamic view of how these axons elongate and find their way to their final target. A major advance in that field came with the use of LSFM on transparent model organisms such as C. elegans or zebrafish (Abu-Siniyeh and Al-Zyoud, 2020; van Krugten et al., 2021). Thanks to the increased imaging speed and low phototoxicity provided by LSFM, transgenic embryos can now be continuously imaged during embryogenesis, providing an unprecedented ability to monitor axonal behaviors and growth cone dynamics over long periods of time in vivo. Changes in growth cone morphology were correlated with guidance decisions made by neurites circumnavigating the nerve ring in C. elegans after imaging embryos every 2 s for 14 h (Wu et al., 2011). Computational tracking of the movements made by ALA mechanosensory neurons and their neurites in elongating embryos further revealed that anterior-posterior neurite outgrowth starts towards the end of elongation and continues after cells reach their final positions (Christensen et al., 2015). In zebrafish, live LSFM was used to visualize branchiomotoneuron axon outgrowth and sensory cranial ganglion projections (Ingold et al., 2015; Zecca et al., 2015). In vivo time-lapse analysis of sensory neurons notably revealed a link between neuronal differentiation and the topography of sensory projections: the site of differentiation prefigures the location where sensory axons enter the hindbrain, whereas the order of differentiation determines the mediolateral topography of the sensory bundles within the hindbrain (Zecca et al., 2015). More recently, imaging of axons and growth cones in the spinal cord of zebrafish embryos in which a subset of newly differentiated neurons were genetically labeled with fluorophores led to the discovery of different modes of elongation: growth cones of axons migrating in the rostrocaudal direction explore their environment in the direction of their motion, whereas growth cones of dorsoventrally aligned axons probe across a larger two-dimensional fan (Liu et al., 2018) (Fig. 1). LSFM has also been coupled with lineage-tracing algorithms and functional ablation approaches to delineate principles of network assembly in C. elegans and discover an essential role for pioneer neurons in strata formation (Moyle et al., 2021).

Along with establishing how the birth, position or migration of neurons influences the spatiotemporal dynamics of axon growth and circuit wiring, LSFM has allowed the discovery of subtle phenotypes in signaling mutants that would have been challenging to reveal with earlier methods. For example, live imaging of cranial motor axon outgrowth in zebrafish embryos uncovered a novel function for the cell-adhesion molecule MDGA2A. In absence of MDGA2A, branchiomotor axons make pathfinding mistakes and exhibit increased collateral branching and defasciculation (axons aberrantly leave from axonal bundles) (Ingold et al., 2015). In the mouse, three-dimensional imaging of immunolabeled axonal tracts in netrin 1 (Ntn1), Dcc and Robo3 mutants unraveled specific and unexpected defects in medial habenula axons forming the fasciculus retroflexus (FR) tract (a tract that relays information from the habenula to the interpeduncular nucleus of the limbic midbrain) (Belle et al., 2014). FR axons surprisingly cross the floor plate but then turn back and coalesce at the midline in Robo3 mutants, which differs from the guidance errors made by other commissural tracts. Conversely, FR axons not only fail to cross the midline but also show additional fasciculation and pathfinding defects in Ntn1 and Dcc mutants, demonstrating distinct, specific functions for these guidance signaling pathways in the habenular system (Belle et al., 2014). As the fine details of circuit wiring can now be resolved, researchers have begun to revisit the role of well-known signaling pathways in brain development. For example, LSFM was recently combined with a novel eye-clearing method to reassess the role of DCC in the visual system (Vigouroux et al., 2020). Retinal axons were found to stall at the optic disk and make pathfinding errors in central visual centers that persist postnatally in the absence of DCC in the retina (Vigouroux et al., 2020). Altogether, by allowing imaging of the nervous system like never before, LSFM offers a detailed and dynamic picture of circuit wiring in vivo that not only extends our understanding of how the brain forms in four dimensions, but also shines new light on the molecular mechanisms governing nervous system assembly. The parallel development of SRM has further revolutionized the study of axons at the mesoscale range, casting a spotlight on how intra-axonal structures and protein dynamics underlie the adaptive response of axons to their environment.

Deciphering how individual axons assemble into circuits requires the ability to resolve sub-axonal structures and signaling complexes at the nanoscale resolution. The recent advent of SRM (see Box 1 for a brief overview of SRM techniques, including SIM and PAINT) has transformed our view of the molecular processes driving axon growth and guidance by enabling the visualization of cell-surface receptors, organelles, actin and microtubule cytoskeletons, and intracellular signaling events below the diffraction limit, in the 10-100 nm range (Fig. 2). SRM has notably enhanced the axial resolution of imaging from about 500 nm (resolution of confocal miscroscopy) to 50-200 nm (see Table 1 for details) (Gustafsson et al., 2008; Hein et al., 2008; Jones et al., 2011), thereby allowing unprecedented observations in three dimensions.

Axons respond to a panoply of attractive and repulsive factors present in their environment to find their way to their post-synaptic target. Their responsiveness to these guidance cues is dictated by the presence of corresponding receptors at the growth cone surface and varies spatially and temporally depending on receptor availability. The distribution of receptors is indeed dynamic, with receptors actively targeted to the plasma membrane by exocytosis, redistributed across the membrane by lateral diffusion, or removed from the plasma membrane by endocytosis or extracellular cleavage (Zang et al., 2021). Although confocal imaging has been instrumental in defining the repertoire of receptors expressed by a given axon at a specific time, the diffraction-limited resolution it achieves does not allow for the characterization of receptors and their dynamics at the molecular level. We do not know, for example, how many receptors must interact with each other to transduce signals or whether they need to be targeted to specialized sub-domains at the plasma membrane to do so. In contrast to confocal microscopy, SRM offers a unique opportunity to address these questions by enabling the precise visualization of single molecular complexes in living cells at the nanoscale resolution. Dual-color stochastic optical reconstruction microscopy (STORM) imaging of the membrane protein caveolin 3 and the ryanodine receptor RyR in cardiac myocytes, for example, reveals distinct localizations for both proteins, demonstrating that their apparent colocalization observed by confocal microscopy is actually caused by optical blurring (Baddeley et al., 2011) (Fig. 2A,A′).

Elegant studies have first employed SRM to dissect the nanoscale organization of receptors at the synapse. Glutamate AMPA receptors, for example, were found to concentrate in sub-synaptic domains of about 70 nm and randomly diffuse between them within seconds by lateral diffusion (Nair et al., 2013). Similarly, metabotropic glutamate receptors mGluR4s and the adhesion molecule neurexin 1 appear organized in small, discrete nanodomains at pre-synaptic active zones (Siddig et al., 2020; Trotter et al., 2019). A similar partitioned organization of receptors and adhesion molecules has more recently been observed along the axon shaft. STORM and stimulated emission depletion (STED) imaging studies have revealed that the endocannabinoid receptor CB1 and the neural cell-adhesion molecule NCAM1 are both distributed into clusters that are evenly spaced along axons, with a periodicity of about 190 nm (Li et al., 2020; Zhou et al., 2019). Interestingly, this distribution coincides with the membrane-associated periodic skeleton (MPS): actin, spectrin and associated molecules generate a braided lattice along axons, with actin forming ring-like structures wrapping around the axonal circumference that are evenly spaced every ∼180 to 190 nm (Vassilopoulos et al., 2019; Xu et al., 2013) (Fig. 2E-E″). The MPS acts as a signaling platform that CB1 and NCAM1 are recruited to upon activation by agonists, leading to the trans-activation of receptor tyrosine kinases and associated downstream signaling (Zhou et al., 2019). As both NCAM1 and CB1 are known to regulate axon guidance and fasciculation (Berghuis et al., 2007; Hansen et al., 2008; Mulder et al., 2008; Saez et al., 2020), it will be interesting to test how their trafficking and compartmentalized signaling contributes to the formation of neural circuits.

Changes in the nanoscale distribution of receptors play a major role in regulating the responsiveness of elongating axons to extracellular factors. A remarkable study illustrated this notion perfectly by tracking the distribution of plexin and Robo receptors at the surface of spinal commissural axons crossing the midline at the floor plate (Pignata et al., 2019). Midline crossing requires axons to change their sensitivity to the floor plate-derived repulsive cues Slit and semaphorin 3B (Zou et al., 2000). Initially insensitive, axons become responsive as they reach the midline, thereby being propelled away from that region as they cross it. STED analysis of axons in spinal cord open book preparations revealed that commissural growth cones alter the presentation of plexin A1 (Plxna1) and Robo1 receptors in a specific spatio-temporal sequence as they approach the midline (Pignata et al., 2019). Plxna1 is delivered to the growth cone surface during navigation through the first half of the floor plate (from the floor plate entry to the midline), whereas Robo1 appears at the plasma membrane only when growth cones navigate through the second half. Interestingly, Plxna1 predominantly accumulates at the front of the growth cone (Fig. 2C,C′), whereas Robo1 concentrates at its rear. The recruitment of receptors to the plasma membrane correlates with changes in growth cone behavior: growth cones slow down after gaining Plxna1 but increase their exploratory behavior along the rostro-caudal axis as Robo1 appears at the surface. Whereas Plxna1 maintains its restricted localization during growth cone pathfinding to the midline, the distribution of Robo1 actively changes and becomes more diffuse as the growth cone reaches the floor plate exit (Pignata et al., 2019). Altogether, these detailed observations underscore that receptors are extremely dynamic and must be delivered to the surface of growth cones at specific times and locations for proper axon navigation and circuit wiring.

Membrane dynamics not only ensure proper trafficking of receptors to and from the plasma membrane but also mediate the response of a growth cone to a cue. Seminal studies using total internal reflection fluorescence microscopy (TIRFM) coupled with calcium uncaging revealed that an imbalance between exocytosis and endocytosis downstream of calcium signaling is required for steering the growth cone in one direction (Tojima et al., 2014). VAMP2-mediated exocytosis exceeds clathrin-mediated endocytosis to ensure proper growth cone turning towards a factor. Conversely, predominant endocytosis mediates repulsion (Tojima et al., 2014). Interestingly, endocytosis at the leading edge of the growth cone was recently found to be clathrin independent (Nozumi et al., 2017). Instead, analysis of vesicular trafficking in growth cones by structural illumination microscopy (SIM) has identified a novel type of endocytosis at the leading edge that relies on endophilin and is linked to F-actin bundling (Fig. 2D,D′). Endophilin-mediated endocytosis removes cholesterol-rich membrane domains such as lipid rafts and appears to sustain growth cone morphology and axon outgrowth (Nozumi et al., 2017). The importance of vesicular release for proper axonal navigation has conversely recently been demonstrated in vivo. Using deconvolution coupled with stepwise optical saturation microscopy (DeSOS) on confocal time-lapse images to achieve super-resolution (Zhang et al., 2019), researchers detected an accumulation of synaptic vesicles in pioneer growth cones of dorsal root ganglia sensory axons as axons crossed the dorsal root entry zone (DREZ) in zebrafish (Nichols and Smith, 2019a). Interestingly, the matrix metalloproteinase MMP14a was found to colocalize with growth cone actin clusters at the DREZ. Preventing vesicular release or inhibiting MMP14a disturbed pioneer axon spinal invasion, suggesting that the delivery of synaptic-like vesicles and MMP14a at the DREZ decision point enables growth cone navigation across the glial limitans. Importantly, vesicular accumulation was found to depend on the formation of actin-based invasive protrusions, demonstrating an essential role for cytoskeletal rearrangements in axon pathfinding (Nichols and Smith, 2019a).

The actin cytoskeleton has long been known to be crucial for proper axon pathfinding (Gomez and Letourneau, 2014). Filamentous actin (F-actin) accumulates at the growth cone and actively polymerizes, depolymerizes and assembles into a mesh or bundles to form peripheral lamellipodia or filopodia, respectively. Actin polymerization and recycling are especially important for growth cone adaptive responses as they provide the forces driving filopodial and lamellipodial explorative protrusions. Recent studies have taken advantage of SRM to discover novel mechanisms regulating actin dynamics in growth cones. For example, tracking of single molecules using photo-activated localization microscopy (PALM) coupled to TIRFM revealed that the core planar cell polarity protein Vangl2 reduces axon elongation by decreasing the mechanical coupling between N-cadherin adhesions and the retrograde actin flow (Dos-Santos Carvalho et al., 2020). F-actin was also found by STED microscopy to depend on the centrosome and the translocation of somatic F-actin from the cell body, as its treadmilling speed significantly decreases after centrosomal disruption (Meka et al., 2019). SRM has further led to the discovery of additional F-actin structures sustaining growth cone behaviors. F-actin-rich puncta were detected by SIM in the central region of growth cones where they colocalize with point-contact and invadosomal proteins such as integrin and cortactin (Santiago-Medina et al., 2015). The resulting apical invadosomal protrusions were shown to dynamically extend from the central domain and to contain proteases such as ADAM17. Inhibiting invadosome formation strongly decreases the extension of motor axons into the peripheral myotome in Xenopus, suggesting that invadosomal proteolytic activity might be required for local extracellular matrix degradation and growth cone advance (Santiago-Medina et al., 2015). Similarly in zebrafish, invadosomes were observed to form in the central domain of pioneer sensory growth cones as they enter the DREZ, and blocking their formation prevents sensory axons from entering the spinal cord (Nichols and Smith, 2019b). Altogether, these studies highlight an important and conserved function of invadosomes in axon navigation in vivo.

Besides new actin arrangements in growth cones, SRM has revealed unanticipated actin structures along axons. For example, observation of actin dynamics using DeSOS super-resolution time-lapse imaging in zebrafish revealed the existence of actin accumulations at the basis of extending neurites named stable base clusters. These clusters form and disappear immediately before the initiation and retraction of neurites, suggesting that their presence dictates neurite formation (Zhang et al., 2019). Along the axon shaft, SRM has revealed the organization of the MPS and its regulation by non-muscle myosin II to ensure axon radial contractility (Costa et al., 2020; Vassilopoulos et al., 2019; Wang et al., 2020; Xu et al., 2013). SRM has also led to the refined characterization of axonal actin waves and trails. Actin waves were initially characterized as growth cone-like membrane protrusions forming at the base of neurites and migrating to their ends at an average speed of ∼2-3 μm min⁻¹ (Ruthel and Banker, 1998). Because of their anterograde transport and eventual fusion with growth cones, actin waves have been proposed to support fast axonal elongation (Flynn et al., 2009; Ruthel and Banker, 1998). However, a recent study using live STED imaging challenged this hypothesis by observing a decrease in neurite elongation speed (to about 0.5 μm min⁻¹) in the presence of actin waves (Mortal et al., 2017). Actin waves appeared to play a role in maintaining growth cone motility rather than controlling neurite elongation. They were also found to have a different nanoscale organization of acto-myosin structures that sustain a distinct mode of movement compared with growth cones (Mortal et al., 2017). In addition to actin waves, focal hotspots of actin spaced at ∼3-4 μm along axons have recently been detected by live STORM imaging (Ganguly et al., 2015). These hotspots were found to serve as actin polymerization sites that generate long filaments named actin trails that extend bidirectionally along the axon shaft. Although their function still remains unknown, actin trails have been proposed to participate in the slow transport of actin in axons and mediate actin enrichment at pre-synaptic boutons.

F-actin works in concert with microtubules to modulate growth cone advance and steering. In fact, both microtubule transport and polymerization play a major role in axon elongation and guidance (Poulain and Sobel, 2010). Yet the respective contribution of microtubule assembly versus transport to axon elongation has remained a controversial debate for decades. Recent SRM studies have provided new insight by visualizing the dynamics of individual microtubules during axon growth. Single-molecule switching nanoscopy conducted on Aplysia californica growth cones revealed that inhibiting microtubule assembly decreases both microtubule extension in the growth cone and the association between the motor protein dynein and microtubules, indicating that microtubule dynamics are actually required for dynein-mediated microtubule translocation (McElmurry et al., 2020). Additional STED and SIM studies have shown that microtubule dynamics are controlled by actin itself in the growth cone. In Xenopus, microtubules and F-actin were found to align in the growth cone periphery to mediate proper growth cone response to guidance cues (Slater et al., 2019). In contrast, growth cones were found to contain myosin II-generated actin arcs that prevent microtubules from extending in the growth cone, thereby promoting axon pausing (Schaefer et al., 2002; Dupraz et al., 2019). The Rho-GTPase RhoA regulates the formation of these actin arcs, thereby controlling the duration of growth and pause phases that dictate the speed of axon elongation (Dupraz et al., 2019). Interestingly, a recent study challenged the importance of actin arcs for regulating axon growth in vivo (Santos et al., 2020). Whereas actin arcs are observed in growth cones extending in two-dimensional substrates, they are lacking in growth cones navigating in three-dimensional collagen or hyaluronic acid matrices. Consequently, unrestrained microtubules extend into the leading edge of growth cones, thereby promoting faster axonal elongation (Santos et al., 2020). Altogether, these results indicate a different mode of axon growth in a three-dimensional environment and highlight the importance of using SRM on more physiological models to define the molecular mechanisms sustaining axon elongation in vivo.

The advent of LSFM and SRM has propelled the study of neuronal circuit wiring into a new era by offering an unprecedented view of the structural and molecular components governing axon development and network assembly. Yet deciphering precisely how a multitude of axons elongating at the same time in vivo interpret their environment in a specific manner to find their post-synaptic targets remains a challenging task. Axons often behave differently in vitro versus in vivo and can navigate over long distances in deep locations that are difficult to image at high resolution. Our ability to ‘see individual trees in detail, while still looking over the whole forest’, as Maglione and Sigrist elegantly said (Maglione and Sigrist, 2013), will require imaging modalities that combine large-volume and deep-tissue imaging with nanoscale resolution and fast acquisition.

High acquisition speed is crucial for visualizing developing axons in a constantly changing environment in vivo. LSFM has achieved millisecond-scale imaging of neuronal activity in zebrafish brains (Farrar et al., 2018; Haslehurst et al., 2018), but it lacks nanoscale spatial resolution. Current studies are addressing acquisition speed by reducing the number of exposures needed for STORM imaging (Gaire et al., 2021) or by combining LSFM with a radial fluctuation Bayesian analysis (Chen et al., 2020). Used to image the intact brain in Drosophila, the latter approach was able to resolve individual axon terminals with a spatial resolution of 70 nm and at speeds that were two orders of magnitude faster than previous single-molecule techniques (Chen et al., 2020). On the other end of the spectrum, SRM has facilitated the detection of molecular complexes at a mesoscale resolution but has mostly been used on cultured cells or fixed tissue sections to overcome the limited penetration of fluorescent signals in live tissue. To circumvent that problem, several groups are now combining STED with two-photon excitation in order to reveal structural details and conduct live-cell imaging in deep brain tissue (Bethge et al., 2013; Ding et al., 2009; Nägerl et al., 2008). Although no single technology has yet been developed to image large volumes in deep tissues at high speed and super-resolution, single molecule-guided Bayesian localization microscopy (SIMBA) and its derivatives offer some promise (Xu et al., 2017). Indeed, SIMBA provides exceptional spatial (50 nm) and temporal (0.15 to 0.2 s) resolutions, and does not require specialized SRM setups necessary for STED, SIM or PALM/STORM, making it more accessible. Notably, open-source plug-ins that run on simple desktop machines are available for SIMBA processing (Li et al., 2020). Last but not least, recent technological developments are poised to transform our analysis of axon behavior by allowing the observation of molecular interactions at high resolution in vivo. For example, two-color STED has just recently been combined with Förster resonance energy transfer (FRET) to image protein-protein interactions at a single-molecule detection level (Szalai et al., 2021). The prospect of combining SRM with optogenetic control of axon guidance (Endo et al., 2016; Harris et al., 2020) for directly manipulating axons and developing circuits in vivo also promises an exciting future in developmental neurosciences.

The authors apologize for omitting any references owing to space restrictions and thank all reviewers for their constructive feedback.