Cytoskeletal organization of axons in vertebrates and invertebrates

Axons are the cable-like processes of neurons that conduct nerve impulses and transmit this information at synapses to other cells. Axons can be up to 1 m long in humans and potentially up to 30 m long in blue whales (Smith, 2009). Axons are delicate structures, subject to enormous mechanical challenges imposed from the outside (e.g., bending and stretching through body movement) and the inside (e.g., axonal transport of large cargoes). In spite of these challenges, most axons cannot be replaced throughout an organism's lifetime, and mammals lose up to 40% of their axon mass toward high age (Calkins, 2013; Marner et al., 2003). Combatting such deterioration requires active maintenance mechanisms, and understanding them will provide new strategies to address the problem of axon loss in aging and disease (Hahn et al., 2019). To this end, we need detailed knowledge of axonal architecture.

An important property to be considered is that neurons and their axons can display substantial variations. For example, axon calibers can vary enormously between distinct neuron types (as detailed below); the need to minimize space and energy demand favors thin fibers, whereas the need to send information at high velocity and/or spike frequencies requires large-diameter fibers (Friede et al., 1984; Perge et al., 2009, 2012). Furthermore, neurons can be of different polarity (multipolar, unipolar, or pseudomonopolar), depending on their animal species of origin and functions as motorneurons, interneurons, or sensory neurons (Fig. 1 A); whether axon architecture needs to be adapted to these different morphologies remains an open question.

Neurons usually display lengthy axon shafts (light blue in Fig. 1 A) with a range of common features, including organelles, continuous networks of smooth ER, and an extensive cytoskeleton (Hahn et al., 2019). The axonal cytoskeleton of invertebrates and vertebrates alike shows conserved features, such as central longitudinal F-actin trails and cortical F-actin rings arranged by spectrins into a periodic pattern (Xu et al., 2013; Leterrier et al., 2017a). Of these, the actin rings seem to have important roles in regulating axonal stiffness (Krieg et al., 2017; Dubey et al., 2020), axon diameter (Costa et al., 2018, 2020) and microtubule (MT) polymerization (Qu et al., 2017). MTs of various lengths are organized into longitudinal parallel bundles that run along axons, often accompanied by intermediate filaments (neurofilaments [NFs]). MTs have a key role as the essential highways for axonal transport, as structural backbones of axons, and as the source of dynamic MTs that can contribute to morphogenetic processes such as axon branching (Hahn et al., 2019). But do MTs perform these functions based on common mechanistic or organizational principles that apply to the wide range of neuron types and axon diameters across the animal kingdom? To which degree can MT functions be replaced by NFs, which are abundant in some organisms and completely absent in others (Peter and Stick, 2015)?

To address these questions, modern superresolution imaging techniques provide us with powerful strategies (Leterrier et al., 2017a), but there are clear limitations. For example, resolving closely spaced MTs in axonal bundles still poses a challenge (Mikhaylova et al., 2015; Tas et al., 2017). Therefore, EM analyses remain pivotal. However, the bulk of existing EM studies (performed mainly on axons of vertebrates, to a lesser degree on invertebrates; see below) dates back to the final three decades of the last century. Those studies were interpreted with the very limited molecular and mechanistic knowledge existing at the time. This review will therefore revisit those publications and interpret classical structural findings in light of state-of-the-art molecular understanding. First, it will focus on axon initial segments (AIS; a specialized zone of the proximal axon in vertebrate motor- and interneurons; Fig. 1 B). It will then take a comparative look at the organization of MTs and NFs in the shafts of small- and large-caliber axons of vertebrate and invertebrate neurons to address whether and how different species of origin or size differences correlate with different concepts of axon architecture.

To uphold the polarity of vertebrate inter- and motorneurons, components specific to the somatodendritic compartment have to be prevented from entering the axon (Fig. 1 Ai). This is thought to involve a staggered filter system including the axon hillock (referred to as the preaxonal exclusion zone [PAEZ]; Farías et al., 2015; details in Fig. 1 B) and the highly specialized AIS (see below); even the remainder of the axon shaft appears to display filter properties, as illustrated by microtubule-associated protein 2 (MAP2), which advances only poorly into the axon shaft if the AIS is abrogated (Jenkins et al., 2015).

The AIS is a specialization of vertebrate inter- and motorneurons that is tens of micrometers long and up to a few micrometers wide (Table S1), characterized by at least two prominent ultrastructural features reported consistently across different neuron types and animal species: parallel MT fascicles and a conspicuous electron-dense undercoat of the plasmalemma (Table S1; Fig. 1 B; and Fig. 2, A–D). According to EM studies of mitral cells in the mouse olfactory bulb, MT fascicles are not fully established until the end of embryogenesis and the dense undercoat not until days after birth (Hinds and Ruffett, 1973).

These AIS-specific structural features correlate with reports of proteins enriched at the AIS, usually appearing after axons have formed; they include the large cortical scaffold proteins AnkG (ankyrin-G) and nonerythrocytic β4-spectrin, filament-forming septins, the MT-binding ring-finger protein TRIM46 (tripartite motif–containing 46), the MT-binding factor EB3 (end-binding protein 3), the dynein regulator Ndel1 (neurodevelopment protein 1 like 1), the actin regulator Mical3 (molecule interacting with CasL), cell adhesion molecules such as neurofascin-186, and voltage-gated sodium channels and potassium channels; many of these components, as well as myosin 2 activity, were shown to contribute to AIS formation and maintenance (Berger et al., 2018; Fréal et al., 2019; Hamdan et al., 2020; Hedstrom et al., 2008; Jenkins et al., 2015; Kuijpers et al., 2016; Leterrier et al., 2017b; Salzer, 2019; Sobotzik et al., 2009). Several of these proteins can be clearly assigned to MT fascicles and/or electron-dense undercoat.

Thus, fasciculated MTs are connected via ∼25-nm-long electron-dense lateral bridges, thus forming sheath-like arrangements (curved arrows in Fig. 3 C) that appear like pearls on a string when analyzed in axonal cross sections (Fig. 2, A–D). The sizes of these fascicles are extremely variable, as illustrated by examples in Fig. 2, A–D that show between 2 and 23 MTs per fascicle (see also Palay et al., 1968). The electron-dense lateral bridges contain and require TRIM46 (Fig. 3 B), which is widely expressed in the nervous system (Harterink et al., 2019; van Beuningen et al., 2015). The MT–MT cross-linker MTCL1 (microtubule cross-linking factor 1) was shown to promote MT fascicle alignment; however, its role is restricted to the AIS of Purkinje cells (Satake et al., 2017), leaving open how MTCL1 functionally relates to TRIM46, which is likewise present in this cell type (van Beuningen et al., 2015).

The structure of the 35-nm-thick electron-dense undercoat was described in a three-layer model (Fig. 3 A; Chan-Palay, 1972) that agrees with various independent transmission EM (TEM) studies (Conradi, 1969; Kosaka, 1980a, 1980b; Nakajima, 1974). In its radial dimension, the TEM model matches well with a model based on superresolution microscopy of AIS-enriched proteins (Fig. 3 B; Leterrier et al., 2015). However, in the longitudinal dimension, both models diverge: the TEM model proposes a repetitive pattern with a periodicity of 60–80 nm (Fig. 3 A; arrowhead in Fig. 5 b of Nakajima, 1974); in contrast, numerous superresolution analyses have built a convincing case for a periodicity of ∼190 nm, which I could not find reflected in any TEM reports (Fig. 3 B; Huang and Rasband, 2018; Jacko et al., 2018; Jones and Svitkina, 2016; Leterrier, 2018; Leterrier et al., 2017a; Schlüter et al., 2019; Sohn et al., 2019). To explain this mismatch, refined EM studies will be required.

50 yr ago, the MT fascicles at the AIS were discussed to be required for “protoplasmic streaming into the axon” and the dense undercoats for the “production of the electrical signal” (Palay et al., 1968). Today we know that AIS specializations help to filter somatodendritic components (preventing them from diffusing or being transported into axons) and establish the spike initiation zone (SIZ), where action potentials are initiated.

First, the AIS serves as a diffusion barrier for nonaxonal lipids and transmembrane proteins (Kobayashi et al., 1992; Nakada et al., 2003; Winckler et al., 1999), potentially requiring components of the dense undercoat such as cortical actin rings, AnkG, and β4-spectrin (Albrecht et al., 2016; Machnicka et al., 2014; reviewed in Huang and Rasband, 2018).

Second, the filtering of transported somatodendritic cargo at the AIS was proposed to involve actin patches or polarized actin networks, which are regulated by Mical3 and mediate the reversal of somato-dendritic cargo transport via the actin-based motor protein myosin-Va (Al-Bassam et al., 2012; Balasanyan et al., 2017; Hamdan et al., 2020; Lewis et al., 2009; Watanabe et al., 2012). This said, other studies have questioned the existence of adequate actin networks at the AIS (Jones et al., 2014; Jones and Svitkina, 2016). Alternatively or in addition, MT-based mechanisms were proposed that involve the selective activation of dynein-mediated retrograde transport through Ndel1 (Huang and Rasband, 2018; Leterrier, 2018; Kuijpers et al., 2016; Ye et al., 2020). Notably, both actin- and MT-based filtering mechanisms seem to involve a tug-of-war between motor proteins that pull in opposite directions, potentially strong enough to cause MT disorganization if thrown out of balance (Kuijpers et al., 2016). Such mechanical forces likely require a more stable architecture, potentially explaining why MTs arrange into tight fascicles at the AIS.

Furthermore, it has been shown that MTs at the AIS are anchored to the axonal surface through binding via EB3 to the tail domains of AnkG (Fig. 3 B; Fréal et al., 2019; Leterrier, 2018; Vassilopoulos et al., 2019); this interaction is essential for MT fasciculation (Sobotzik et al., 2009). Furthermore, the dynein regulator Ndel1 binds to the AnkG tail (asterisks in Fig. 3 B), and this interaction is essential for AIS filter functions (Ye et al., 2020). In agreement with these functional studies, EM analyses suggest that a fraction of MTs are in contact with the undercoat (Fig. 3 C). However, many MT fascicles are positioned centrally in the axon, clearly out of reach of AnkG (Fig. 3 D). This raises the question as to whether only peripheral MT fascicles are involved in transport filtering and whether central MT fascicles drift freely in the axoplasm or are linked up via inconspicuous linkers that ultimately depend on AnkG.

Third, the AIS is a SIZ, meaning it triggers action potentials as the integrated outcome of supra-threshold depolarizations propagating from the somatodendritic compartment (Kole and Stuart, 2012). This function requires a well-balanced enrichment of voltage-gated sodium and potassium channels at the AIS (NaV and Kv in Fig. 3 B), which are likewise anchored via AnkG (Akin et al., 2015; Hefting et al., 2020). Importantly, neurons can fine-tune their excitability by changing the length of the AIS, the composition and posttranslational modification of ion channels, or the distance from the soma through mechanisms that involve calcium/calcineurin, MT stabilization, and myosin II (Berger et al., 2018; Bolós et al., 2019; Evans et al., 2015, 2017; Grubb and Burrone, 2010; Hatch et al., 2017; Kuba et al., 2010, 2014, 2015).

In various neuron types, the excitability at the AIS is also regulated by synapses in their majority inhibitory input synapses (Fig. 2, C and E; and Table S1). Their formation can be mediated by AnkG-dependent recruitment of the adhesion factor neurofascin-186 (Nfas in Fig. 3 B; Ango et al., 2004; Kriebel et al., 2011) and is regulated by synaptic activity (Pan-Vazquez et al., 2020). As listed in Table S1, many AIS descriptions highlight the occurrence of varicose spines of variable lengths and shapes (Fig. 2, E and F; see also Sobotzik et al., 2009) which seem to lack MTs and sometimes harbor more than one synapse. Beneath input synapses or spines the dense undercoat is consistently reported to be interrupted (Fig. 2 C), and many authors mention the presence of cisternal organelles (Table S1) that bear some resemblance with the spine apparatus (i.e., stacks of ER required for synaptic plasticity; Deller et al., 2003).

In accordance with the polarity of vertebrate motor- and interneurons, the AIS is both a filter zone and a SIZ (Fig. 1 A, i). A similar combined compartment might be expected in the initial axon segment of arthropod sensory neurons (which have similar polarity; Fig. 1 A, iv), but not in vertebrate sensory neurons or arthropod motor- and interneurons (Fig. 1 A, ii and iii). Unfortunately, only few data are available for those neuron classes (legend of Fig. 1 A; Smarandache-Wellmann, 2016).

For example, sensory neurons of vertebrates (e.g., dorsal root ganglion [DRG] or trigeminal ganglion neurons) form a stem axon that then bifurcates into a central and peripheral axon arm (Fig. 1 A, ii; Nascimento et al., 2018). Stem axons display fasciculated MTs comparable to the AIS, whereas the electron-dense undercoat is absent (Fig. 2 G; Bird and Lieberman, 1976; Ishikawa et al., 1980; Peach, 1975; Zelená, 1971). Only one report mentions an extremely short (2–4 µm) stretch of membrane with an electron-dense undercoat (Zenker and Högl, 1976).

Consistent with the presence of MT fascicles and similar to the AIS, the stem axon displays TRIM46 and, proximal to it, MAP2 (Fig. 1 A, i versus Fig. 1 A, ii; Gumy et al., 2017; van Beuningen et al., 2015). These proteins may represent a staggered PAEZ/AIS-like filter system (see above), since MAP2 is required to block somatic organelles from entering DRG axons; in addition, MAP2 selectively activates kinesin-mediated transport, in line with a recently proposed “MAP code” for the regulation of polarized MT-based traffic (Gumy et al., 2017; Monroy et al., 2020); these roles of MAP2 demonstrated at stem axons may similarly contribute to PAEZ filter functions at the axon hillock. Outside the stem axon, vertebrate sensory neurons likely have further filter zones (Fig. 1 A, ii), but little is known at present.

Consistent with the lack of the electron-dense undercoat (Fig. 2 F), stem axons do not show enrichment of AnkG, β4-spectrin, or sodium channels (Fig. 1 A, i versus Fig. 1 A, ii; Gumy et al., 2017). This suggests that MT fascicles can form independently of AnkG and that stem axons have no SIZ functionality. Instead, electrophysiological studies have shown that the SIZ of DRGs is located in the peripheral axon, slightly removed from the dendritic compartment (Fig. 1 A, ii); similar to the AIS, this peripheral SIZ is dynamic and moves closer to the dendrites upon sensitization (Goldstein et al., 2019). Whether it displays AnkG expression and/or an electron-dense undercoat remains to be seen.

Nodes of Ranvier (NoRs) are the small unmyelinated segments along myelinated axons required for saltatory conduction (Fig. 1 A, i). NoRs display AIS-like undercoats (15–25 nm thick, 5–10 nm separated from plasmalemma) but no obvious MT fasciculation (Fig. 2 E; Conradi and Skoglund, 1969; Waxman et al., 1972; Elfvin, 1961; Kosaka, 1980b; Peters, 1966; Reles and Friede, 1991; Tsukita and Ishikawa, 1981; p. 104 and 260ff. in Peters et al., 1991). Only the most proximal NoRs in DRG neurons seem to display MT fascicles, potentially due to spillover of cross-linkers from the stem axon (Nakazawa and Ishikawa, 1995). Ethanolic phosphotungstic acid (an anionic heavy metal complex known to label basic proteins at synaptic densities; Jones, 1993) was reported to label the undercoats at both NoR and AIS, suggesting molecular commonalities (Sloper and Powell, 1973, 1979). We now know that NoRs lack TRIM-46, consistent with the absence of MT fasciculation (van Beuningen et al., 2015) but express characteristic AIS proteins consistent with the dense undercoat and excitability of NoRs, including AnkG, β4-spectrin, neurofascin-186, Nr-CAM, and high levels of voltage-gated sodium channel (Fig. 1 A, i; Kordeli et al., 1995; Dzhashiashvili et al., 2007; Yang et al., 2007; Liu et al., 2020). A further site where AIS-like undercoats were shown to correlate with the expression of at least AnkG are the “hot spots” in dendrites of the rat olfactory bulb (Kosaka et al., 2008).

Taken together, key features coinciding at the AIS can occur in separation in other neuron types or compartments. Therefore, using the term AIS for neurons other than vertebrate motor- or interneurons might be misleading; filter zones, diffusion barriers, or SIZs are expected to be distributed differently in neurons of other polarity (Fig. 1 A), and whether their molecular mechanisms match those at the AIS remains to be seen.

The axon shaft is the longest neuronal compartment and can display enormous structural variations (Table S2). With regard to the length of axonal MTs, it was originally believed that all axonal MTs extend from the soma. This was supported by studies in rat sensory and motor fibers showing that MT numbers of the main axon are equal to the added numbers of its daughter neurites, as if MTs at branch points merely decide whether to continue into one or the other branch (e.g., Weiss and Mayr, 1971a). This view was disputed by other findings: for example, in certain rat and crayfish motorneurons, the sum of cross-sectional areas of daughter branches was found to increase up to ∼10-fold, accompanied by substantial increases in MT counts (Nadelhaft, 1974; Zenker and Hohberg, 1973). This clearly indicated that new MTs arise within the axon, and we now know that nucleation, the seeding of new MTs, occurs in the axon shaft (Baas and Ahmad, 1992; Sánchez-Huertas et al., 2016; Stiess et al., 2010). Anchoring mechanisms that orient nucleation events toward the axon tip (Cunha-Ferreira et al., 2018) ensure that MTs have a consistent parallel polarity in axons (Baas et al., 1988; Baas and Lin, 2011; Yau et al., 2016).

Little data are available describing the actual lengths of MTs in axons. From serial TEM analyses of fairly short axon fragments, a range of MT lengths was estimated as follows: 430 µm/97 µm (frog, olfactory axons in the olfactory nerve/olfactory bulb; Burton, 1987), 108 µm (rat, cultured sensory nerves; Bray and Bunge, 1981), 513–757 µm/374 µm (mouse, internodal/nodal regions of sensory saphenous nerve; Tsukita and Ishikawa, 1981) and 9–23 µm/∼5 µm/∼6.5 µm (Caenorhabditis elegans, mechanosensory axons/axons of GABAergic dorsal D-type neurons/other axons; Chalfie and Thomson, 1979; Kurup et al., 2015). In agreement with the EM data for C. elegans, advanced live imaging strategies reported ∼6-µm-long MTs in axons of DA9 motorneurons (Yogev et al., 2016). Serial TEM analyses in both C. elegans and rat also reported the direct observation of very short MTs (1–4 µm), which might represent MT fragments undergoing transport (Baas, 2002; Baas et al., 2006). Notably, developing rat hippocampal neurons in culture overwhelmingly showed short fragments (range, 0.05–40 µm; average, 4 ± 5 µm; Yu and Baas, 1994), which might reflect high MT dynamics at the axon growth stage.

The length of MTs potentially impacts axonal transport dynamics (Yogev et al., 2016), but likely more important is the number and density of MTs, which, according to mathematical modeling, correlates with transport volume (Wortman et al., 2014). This has been most convincingly studied in DRG neurons (Fig. 1 A, ii), where the peripheral axon branch has a larger caliber, displays more MTs, and has a higher transport rate than is observed in the central branch of the same neurons (reviewed in Nascimento et al., 2018).

Across the animal kingdom, axon diameters were reported to be in the range of 0.1 to 30 µm in vertebrates, 0.1 to >100 µm in arthropods (giant axons of some spiders, crustaceans, and cockroaches; Castel et al., 1976; Foelix and Troyer, 1980; Nadelhaft, 1974), and up to a millimeter in squid (as the result of axonal fusion of 300–1,500 neurons; Young, 1936; Brown and Lasek, 1990). For axons of vertebrates and arthropods, MT numbers per profile area range between 1 and >250,000 MTs and their densities between 2 and ∼100 MTs/µm² (Table S2), with a reported outliers far above 100 (Fadić et al., 1985).

In light of these enormous ranges, the question arises as to whether the architectural challenges posed by different axon diameters require specific solutions or can be accommodated by common organizational concepts.

Small-caliber axons (<300 nm diameter) can be found in most species (Fig. 4, A, B, G, and H), for example the olfactory bulb or the molecular layer of the cerebellum of vertebrates (Andres, 1965a; Harada et al., 1994; Kreutzberg and Gross, 1977; Peters et al., 1991), peripheral nerves and the ventral nerve cord of insects (Egger et al., 1997; Stephan et al., 2015), and nerve tracts of C. elegans (Chalfie and Thomson, 1979; Chalfie and Thomson, 1982). In these examples, MTs are distributed across the axonal profile; they are spaced from each other but appear only loosely cross-linked. Accordingly, conventional TEM strategies do not reveal conspicuous MT–MT connections as observed at the AIS. However, linkers clearly exist and become visible upon use of other strategies. For example, lanthanum is a trivalent cation that penetrates cell membranes if applied during fixation and severely enhances contrast of many subcellular structures upon TEM analysis (Leeson and Higgs, 1982; Shaklai and Tavassoli, 1982). In lanthanum-stained axons, MTs display irregular coating and fuzzy connections become visible between them (Burton and Fernandez, 1973; Lane and Treherne, 1970; Ochs and Burton, 1980). In classical freeze-etch preparations, nerves were detergent extracted, quick frozen, mechanically fractured to expose the inside, etched via freeze drying, and then rotary shadowed with platinum/carbon to produce a replica for analysis. Freeze-etch preparations clearly revealed 4- to 6-nm-thick MT–MT cross-bridges (Chen et al., 1992; Harada et al., 1994; Hirokawa, 1982, 1991).

Although nearest-neighbor MT–MT distances are not constant, their means show reproducible values that differ between neuron types and species (e.g., Fig. 4, A, F, and G); MTs are spaced on average ∼22 nm in parallel fibers of the mouse cerebellum (Chen et al., 1992), ∼110 nm in Drosophila (Egger et al., 1997; Stephan et al., 2015), and 9.8 nm in C. elegans (Krieg et al., 2017). In all of these cases, loss-of-function studies have revealed specific MAPs as MT–MT cross-linker candidates; loss of the Drosophila MAP1b homologue Futsch caused a reduction in nearest-neighbor values from ∼110 to ∼70 nm (Bettencourt da Cruz et al., 2005; Stephan et al., 2015), loss of ptl-1 (the C. elegans tau homologue) caused a reduction from 9.8 to 6.7 nm (Krieg et al., 2017), and loss of murine Mapt/tau abolished the majority of cross-bridges in freeze-etch preparations (Harada et al., 1994).

Differences in MT spacing become meaningful when comparing the longest isoforms of these three MAPs (all data taken from ensembl.org): ptl-1 (458 aa) is shorter than tau (749 aa), whereas Futsch is dramatically longer (5,495 aa). Futsch has a large repeat-rich central domain and its terminal regions show homology with the N- and C-terminal parts of human MAP1B, excluding however the two designated MT-binding domains (Fig. S1; Hummel et al., 2000; Villarroel-Campos and Gonzalez-Billault, 2014). It remains therefore to be seen whether Futsch achieves direct or indirect MT association through its terminal homology regions, which could mean that a single molecule could act as a MT–MT cross-linker.

In contrast, tau proteins binding to adjacent MTs were proposed to dimerize via their N-termini (Rosenberg et al., 2008), and the N-termini are clearly longer in tau (∼450 aa) than in ptl-1 (∼300 aa). A more recent model for tau-mediated MT–MT cross-linkage is less static and allows tau to diffuse (Méphon-Gaspard et al., 2016), consistent with observations that tau has a very short dwell time on MTs (Janning et al., 2014; Samsonov et al., 2004). Such loose binding dynamics might be one explanation for why MTs are not firmly interlinked into fascicles or hexagonal patterns but are free to detach and “diffuse” into empty areas of the axonal profile. This said, at high expression levels MAPs can very well induce hexagonal MT patterns (Chen et al., 1992), suggesting that MAP levels are well controlled in axons.

Surprisingly, the loss of specific MAPs (see above) did not cause a complete collapse of MTs onto each other. This may be explained by repulsion due to the high negative surface charge of MT lattices, but it more likely occurs through the expression of potentially redundant MAPs, as reviewed elsewhere (Bodakuntla et al., 2019; Burgoyne, 1991; Hahn et al., 2019; Hirokawa, 1991; Lee and Brandt, 1992; Matus, 1991). As a further complication, MAPs may change their affinities for MTs as a function of posttranslational tubulin modifications or tubulin isotypes. For example, polyglutamylation regulates binding of tau to MTs (Boucher et al., 1994). Furthermore, the average MT–MT distance of 9.8 nm in C. elegans (see above) was measured in mechanosensory neurons expressing mec-7 β-tubulin; upon mec7 deficiency, the same mechanosensory axons form MTs of smaller diameter, which display a reduced spacing of only 4 nm (Chalfie and Thomson, 1982; Krieg et al., 2017), suggesting that MTs with different tubulin isotypes are cross-linked differently.

The organizational principle described for small-caliber axons appears to be maintained in large-diameter axons of crustaceans, insects, or spiders; MTs are distributed across the axon profile and, where data are given, nearest-neighbor distances are typically 100–200 nm (Fig. 4, G and H; Bettencourt da Cruz et al., 2005; Burton and Fernandez, 1973; Egger et al., 1997; Foelix and Hebets, 2001; Foelix and Troyer, 1980; King and Wyman, 1980; Lane and Abbott, 1975; Nadelhaft, 1974; Smith et al., 1977; Stephan et al., 2015; Viancour et al., 1987; Warren, 1984).

In crustaceans, lanthanum staining revealed “axoplasmic matrix proteins” in the form of 5- to 10-nm-thick MT–MT cross-bridges (Burton and Fernandez, 1973; Ochs and Burton, 1980). Upon extraction with glycerol-sucrose-DMSO solution, MT–MT distances were seen to collapse, and the extracted protein was of high weight and heat stable, as is typical of MAPs (Warren, 1984). An anti-NF antibody (clone NN18) stained the cross-bridges and detected a 600-kD protein with homology to MAP1B (Weaver and Viancour, 1991; Weaver and Viancour, 1992; but see Corrêa et al., 2004).

These data match very well with those for the Drosophila Futsch protein: like observed in crustaceans, Futsch spaces MTs at ∼110 nm (see above; Stephan et al., 2015), displays MAP1B homology, and its central domain is composed of 60 repeats with intermediate filament homology (Hummel et al., 2000). In support of this notion, arthropod genomes, including crayfish (asterisk in Fig. S2), tend to contain Futsch-like genes, which display homology with Futsch in exactly those N- and C-terminal regions that are also homologous to human MAP1B (Fig. S1).

In addition to MT–MT linkers, also cross-bridges from MTs to the cortex were reported in crayfish (Burton and Fernandez, 1973). Extrapolating from work in Drosophila, these might be large ankyrins (Anks). Thus, long isoforms of cortex-associated Ank2 join forces with Futsch to form MT–MT networks that help to expand and maintain axon calibers in Drosophila (Stephan et al., 2015). Also in C. elegans, large Anks seem to anchor MTs along the axon shaft (He et al., 2020), and large AnkB isoforms appear to perform similar functions along axon shafts in mouse (Yang et al., 2019), consistent with reports of ∼100-nm-long filaments linking the plasmalemma to MTs in freeze-etch preparations of frog axons (Hirokawa, 1982). For a discussion of further candidate factors involved in surface anchorage of MTs, see Hahn et al. (2019).

Taken together, different arthropod species seem to use a consistent strategy of MT cross-linkage across a wide range of axon diameters (∼0.1 to ∼60 µm). Although Futsch-like proteins seem the most prominent players, further MAPs such as Tau or Ensconsin, or even type 6 kinesins, might contribute to these networks (Hahn et al., 2016; Monroy et al., 2018).

A major difference between axons of arthropods and vertebrates is the presence of NFs. Cytoplasmic intermediate filaments are clearly absent in arthropods (Peter and Stick, 2015), whereas NFs are highly abundant in large-caliber axons of vertebrates where they are spaced out and distributed in a similar manner as described for MTs (details below; Fig. 4, D and E; Hirokawa, 1991; Peters et al., 1991).

NFs of vertebrates are ~10 nm thick, highly stretch-resistant intermediate filaments composed of a chain of hetero-oligomeric subunits (Fig. S2). Each subunit contains a mix of type IV proteins encoded by different genes (Fig. S2 A; Herrmann and Aebi, 2016; Kornreich et al., 2015; Perrot et al., 2008), usually including a lightweight (α-internexin, 66 kD; or NF-L, 62 kD), medium-weight (NF-M, 103 kD), and heavyweight (NF-H, 117 kD) protein, from now referred to as α, L, M, and H. These proteins can coassemble with type III peripherin (Yuan et al., 2012). In small unmyelinated fibers, peripherin even seems to be the predominant NF component (Larivière et al., 2002). However, in large myelinated fibers, neither loss nor gain of peripherin has revealed major impacts on NF formation, although long-term overexpression triggers amyotrophic lateral sclerosis–like axonopathy (Beaulieu et al., 1999; Larivière et al., 2002).

NFs promote large axon diameters, as is clearly demonstrated by axon-caliber reduction in NF-deficient mouse and quail models or human patients (Eyer and Peterson, 1994; Sakaguchi et al., 1993; Yamasaki et al., 1991; Yum et al., 2009). H, M, L, and α seem to contribute to different degrees; calibers of large axons were strongly reduced and NFs virtually absent in L^−/− mutant mice, quail, and humans, whereas both parameters were only moderately affected in M^−/− and marginally in H^−/− and α^−/− mutant mice (Beaulieu et al., 1999; Elder et al., 1998; Jacomy et al., 1999; Levavasseur et al., 1999; Ohara et al., 1993; Yamasaki et al., 1992; Yum et al., 2009; Zhu et al., 1997, 1998). However, masked redundant roles of H and α became apparent in H^−/−M^−/− or α^−/−L^−/− double-mutant mice, which displayed complete or next to complete NF depletion (Jacomy et al., 1999; Levavasseur et al., 1999). Importantly, NF proteins have to be present in proportional amounts. Overexpression of M or H (but not L) caused axon-caliber reduction, which was even further enhanced upon MH coexpression; in contrast, LM or LH coexpressions caused caliber increases (Xu et al., 1996).

Caliber enlargement can also be achieved by applying the neurotoxin β,β'-iminodipropionitrile: in treated axons, proximal segments enlarge due to NF accumulation; this crowding of filaments does not affect NF-NF spacing (39–46 nm; Parhad et al., 1987), suggesting that NF networks have the physical ability to drive axon enlargement. Comparable NF-driven axon caliber expansions are also observed during giant axonal neuropathy, primarily in internode regions of myelinated fibers (Asbury et al., 1972; Berg et al., 1972).

NF–NF spacing depends on the tail domains of NF proteins (Fig. S2). The tail domains differ in lengths (H, 679 aa; M, 514 aa; L, 158 aa; and α, 91 aa) and stick out from NFs like the hairs on a pipe brush (Fig. S2; Kornreich et al., 2015; Stevens and Hoh, 2010). Freeze-etch analyses of axons revealed 20- to 60-nm-long and 4- to 6-nm-thick NF–NF cross-bridges with a periodicity of 20–100 nm and containing at least the tail domain of H (Hirokawa, 1982; Hirokawa et al., 1984; Schnapp and Reese, 1982). TEM analyses confirmed the existence of ∼45-nm-long cross-bridges in myelinated axons, whereas they only measured ∼30 nm at the NoR or in dendrites, which both lack myelin (Hsieh et al., 1994; Wuerker and Palay, 1969). A very similar reduction to 30 nm was observed in myelinated axons lacking the myelin component myelin-associated glycoprotein (Yin et al., 1998). This led to the hypothesis that myelin-associated glycoprotein–dependent phosphorylation of M/H tails (Fig. S2) at their ∼7/∼51 KSP motifs triggers cross-bridge extension, hence the increased NF spacing in myelinated axons (Nixon et al., 1994; Garcia et al., 2003; Kumar and Hoh, 2004 and references therein).

Consistent with the knockout data, systematic tail domain deletions (^tailΔ) in transgenic mice revealed a reduction in cross-bridge numbers in axons of H^tailΔ or M^tailΔ mutant mice, but only for M^tailΔ did this result in narrower axons and slightly shorter NF spacing (39 nm); in M^tailΔ H^tailΔ double-mutant mice NF–NF side bridges were absent and axon diameters were further reduced (Garcia et al., 2003; Rao et al., 2002), lending further support to the idea of redundancy between M and H. Surprisingly, M^tailΔ H^tailΔ filaments still ran in parallel to the axon axis and maintained a spacing of ∼30 nm, likely mediated by the shorter tail domains of L and/or α (Kornreich et al., 2015). Consistent with these findings, application of osmotic shock induced by high sucrose in axons or polyethylene glycol in vitro (likely abolishing charge-dependent extensions in particular of H and M tails), caused a severe reduction in NF spacing, but not a total collapse (Friede and Samorajski, 1970; Kornreich et al., 2015).

Therefore, in contrast to modular MT–MT coupling, which depends on separately encoded MAPs, NF–NF linkage and spacing is a function of the filaments themselves and can be regulated through phosphorylation. It appears that their cross-linking affinities are low enough so that NF–NF networks can undergo dynamic changes, as was similarly discussed for MT–MT networks.

NFs tend to be sparse in small axons. However, with increasing diameters, NF numbers rise dramatically (much faster than MT numbers) and may eventually outnumber MTs by an order of magnitude (Table S2; Fig. 4, A–E; Alfei et al., 1991; Alvarez et al., 1982; Malbouisson et al., 1985; Pannese et al., 1981). This trend applies to myelinated and unmyelinated axons (Friede and Samorajski, 1970; Smith et al., 1970), although some exceptional reports find constant MT densities across different caliber ranges (Elder et al., 1998; Zenker and Hohberg, 1973), and NF/MT ratios in the extraordinarily large giant axons of squid seem to be as low as 2:1 (Fath and Lasek, 1988). However, in general, MT densities in large-caliber axons of vertebrates tend to be lower than in arthropods due to the presence/absence of NFs; for example, MT densities in axons of lamprey are approximately an eighth compared with densities in cockroach axons of similar diameter (Table S2).

Images and descriptions of small versus large caliber axons across vertebrate species suggest fundamental differences in MT distribution (Nakazawa and Ishikawa, 1995; Peters et al., 1991; Raine et al., 1971; Tanner et al., 1998; Tsukita and Ishikawa, 1981; and references listed in Table S2). In small fibers (Fig. 4, A–C), MT–MT and NF–NF networks tend to segregate, which may be a function of their composition (e.g., predominance of peripherin; see above) and/or of their numbers. With increasing NF numbers, there is a clear impact on MT organization (Fig. 4, D and E) in that patches of MT–MT networks become rarer and smaller, and often single MTs seem to “dissolve” into the surrounding “sea” of NF–NF networks, thus spreading out at low densities across axonal profiles.

To act as a “solvent,” NFs would have to be able to cross-link to MTs, and this seems to be the case: in nonneuronal cells, the cell-junctional scaffold protein plectin was described as a cross-linker between MTs and the intermediate filament vimentin (Svitkina et al., 1996); in axons, MT–NF cross-bridges have also been described (Hirokawa, 1991; Rice et al., 1980; Schnapp and Reese, 1982). Such cross-bridges were suggested to be MAPs or MAP-like factors (e.g., tau, MAP2, or MAP6) or even motor proteins (Leterrier et al., 1982; Miyata et al., 1986; Perrot et al., 2008). Unlike the regular side-arms of NFs, MT–NF linkers in axons were described to display random periodicity, and their shapes were often bifurcating with a length of 54.7 nm, which was distinct from the spacing of MT–MT (49.6 nm) or NF–NF (40.61 nm) cross-bridges (Monaco et al., 1989; Tsukita et al., 1982).

NFs are proposed to form “hydrogel networks” (Kornreich et al., 2015) and to behave like “inert ‘molecules’ in a dilute solution” that “are relatively free to move apart from each other and to fill the available space within the axon” (Price et al., 1988). This would mean that also MTs dissolved in these networks are fairly free to drift, ideal to accommodate dynamic space requirements, for example when voluminous cargoes are transported through an axonal region. In support of this idea, serial TEM analyses of axons reveal loose parallel arrangements, with MTs changing their relative positions in consecutive axon segments (Fig. 1 A; Bray and Bunge, 1981; Tsukita and Ishikawa, 1981). Furthermore, MTs often cluster around mitochondria (Friede and Samorajski, 1970; Hirokawa, 1982; Malbouisson et al., 1985; Pannese et al., 1981; Raine et al., 1971; Smith et al., 1975, 1977; Tsukita and Ishikawa, 1981). This suggests that MTs can be attracted toward mitochondria, likely mediated by MT–mitochondrial linkers that have a higher affinity than MT–MT or MT–NF linkers. These could be MT-associated motor protein complexes for mitochondria that are being transported (Hirokawa et al., 2010) or the cross-linker protein syntaphilin known to dock or anchor stationary mitochondria to MTs (Kang et al., 2008; Lin et al., 2017). The same logic may apply to occasional groups of narrowly cross-bridged MT fascicles positioned within NF-rich axon shafts of lamprey neurons (Smith et al., 1970, 1975).

Surprisingly, removal of the highly abundant NF–NF networks from axons is not a lethal condition, although up to 20% of axons can be lost, and axon calibers and their conduction velocities are significantly reduced (Beaulieu et al., 1999; Bocquet et al., 2009; Eyer and Peterson, 1994; Jacomy et al., 1999; Križ et al., 2000; Sakaguchi et al., 1993; Yamasaki et al., 1991; Yum et al., 2009; Zhu et al., 1997). In the absence of NFs, the densities of MTs increase up to 10-fold, but they are well organized and look very similar to those of small-caliber vertebrate axons or axons of arthropod neurons; MTs are distributed across the axonal profile and seem to maintain an average nearest-neighbor spacing from each other (Fig. 4 I versus Fig. 4, A–C and F–H).

In lower motorneurons of rat, MAPs localize to MTs scattered throughout the axonal profile (Papasozomenos et al., 1985). This suggests that the typical MT-regulating machinery is readily available so that MTs can interact in meaningful ways even when dissolved into NF–NF networks. Such a scenario could explain why loss of function of the MT-binding and -regulating spectraplakin dystonin causes severe disorganization of axonal MTs regardless of whether NFs are present or absent (Bernier and Kothary, 1998; Dalpe et al., 1998; Eyer et al., 1998; Voelzmann et al., 2017) and that even loss of its fly homologue Short stop causes the same phenotype (Alves-Silva et al., 2012; Sánchez-Soriano et al., 2009).

This said, NF depletion might impact on MTs. For example, NFs were reported to regulate MT stability and/or polymerization (Bocquet et al., 2009; Kurup et al., 2018 and references therein). This might explain why the number of MTs per axon was doubled in H^−/− single-mutant mice, although NF numbers and axon diameters were hardly affected in these animals (Jacomy et al., 1999; Zhu et al., 1998). In contrast, there seems to be no up-regulation of MT numbers in NF-depleted quiver mutant quail axons when calculating increases in MT density against the reduction in cross-sectional area (Yamasaki et al., 1991).

Regardless of whether MT numbers change, NF-depletion provides unique opportunities to study the “default” MT–MT networks in large caliber axons. For example, nearest-neighbor relationships of MTs are impossible to assess in larger axons where MTs are dissolved into NF–NF networks. Depletion of NFs would therefore provide a condition where true spacing could be assessed, thus providing improved strategies to study MAP functions in these neurons.

An efficient experimental strategy to generate NF-free MT–MT networks is provided by the injection of the toxins 2,5-hexanedione or β,β'-iminodipropionitrile. Apart from affecting the proximodistal NF distribution in axons (see above), these toxins trigger a dramatic segregation of the axonal cytoskeleton: within hours or days, a core of MT–MT networks forms, surrounded by a halo of NF–NF networks (Fig. 4 J; Griffin et al., 1983a, 1983b; Papasozomenos et al., 1981; Parhad et al., 1987). Importantly, the central MT–MT networks were shown to form MAP1A/B immunoreactive cross-bridges in sciatic nerve (Hirokawa et al., 1985) and overlap with MAP staining in the ventral root (Papasozomenos et al., 1985), suggesting that MTs brought their MT-regulating machinery into this core when segregating away from NFs.

The underlying mechanisms of this segregation process remain hypothetical. A mathematical model has been proposed stating that cytoskeletal segregation is caused by a reduction in slow MT-based transport of NFs, thus shifting the bias toward other transported cargoes that may cross-link MTs and zip them together (Xue et al., 2015). Similar mechanisms might explain comparable segregations that occur in a number of axonal pathologies, although with certain variations; for example, in giant axonal neuropathy, MT–MT networks form not at the core but in the axon periphery (Asbury et al., 1972; Berg et al., 1972; further references in Xue et al., 2015).

This review aimed to understand concepts of axonal architecture by integrating knowledge from >50 yr of study across a range of animal models, including arthropods that have developed their axonal organization in the absence of NFs. From this, I propose that loose MT–MT networks are a fundamental common element of axons. In these networks, MTs are spaced out and only loosely connected by MAPs, thus allowing for flexible rearrangement, for example to permit and support the passing through of large transported cargoes. The main exception to this rule are fasciculated MTs at the AIS or in stem axons where tight coupling might represent a structural feature to resist the mechanical forces caused by a tug-of-war between antagonistic motor activities.

As discussed elsewhere (Hahn et al., 2019), essential architectural properties of axons depend on stable MTs and a stiff cortex. In contrast, the addition of NF–NF networks appears to have its major purpose in expanding axon calibers, thus improving conduction velocity and spike frequency (Friede et al., 1984; Perge et al., 2009, 2012). However, NFs seem to make astonishingly little contribution to axonal integrity and sustainability. NFs might improve elastic properties of axons, but arthropods clearly demonstrate that caliber enlargements can be achieved solely through increased MT numbers. In large-caliber axons of arthropods, most MTs seem to serve a mere structural purpose, with only a specialized subset used for axonal transport (Miller et al., 1987). This might explain why it was possible during evolution to replace a fraction of the axonal MTs by NFs in other animal groups.

If MT networks are indeed the key element of the longitudinal axonal cytoskeleton, it makes sense that MT-regulating mechanisms are functional in small- and large-caliber axons of vertebrates alike. Accordingly, MTs seem to be equipped with their regulatory toolbox even when submerged into the vast NF–NF networks, enabling them to interact in meaningful ways and to rapidly form intact MT–MT networks when cytoskeletal segregation occurs after 2,5-hexanedione or β,β'-iminodipropionitrile application.

I hope that the descriptions throughout this review have made clear how much information and understanding can be obtained from the retrospective study of classical EM work, which has become easily accessible thanks to the efforts of most journals to systematically digitize their back issues. There is wealth of information for many other contemporary lines of investigation, such as detailed descriptions of MT behaviors during axon branching (Yu et al., 1994), of the ultrastructure of growth cones, dendrites, and synapses (Peters et al., 1991), or of curious phenomena such as impressive packages of ER-derived tubular structures found in certain axons (Andres, 1965b; Peters et al., 1991). Axonal MTs of species as diverse as cockroach, lamprey, frog, toad, chick, mouse, and rats were reported to contain luminal material in form of a ∼4-nm-thick central dot or filament (Andres, 1965a; Burton, 1984, 1987; Gonatas and Robbins, 1965; Lane and Treherne, 1970; Nixon et al., 1994; Rodríguez Echandía et al., 1968; Smith et al., 1970, 1975; Wuerker and Palay, 1969), matching recent reports of MTs with incorporated actin filaments or MAP6/stable tubule-only peptide (Cuveillier et al., 2020; Paul et al., 2019,Preprint), of which the latter could help to explain long-term cold resistance of axonal MTs observed in vivo (Delphin et al., 2012; Pannese et al., 1982; but see also Song et al., 2013). Furthermore, classical studies revealed the presence of post-translationally modified subdomains in MT lattices (Baas and Ahmad, 1992; Baas and Black, 1990; Baas and Joshi, 1992), which start finding their explanations in current models of MT stability (Baas et al., 2016).

In conclusion, I hope that this review will help to build bridges to the past to capitalize on that valuable treasure trove that can complement and often even instruct our work with modern techniques in meaningful ways. But it should also have become clear that existing descriptions only scratch the surface of what we need to understand. Classical EM techniques had clear limitations with respect to efficiency and the use of fixation and staining methods that might not always represent axonal structure in all accuracy. New techniques such as large-volume EM or cryo-electron tomography (Kubota et al., 2018; Xu et al., 2017; Chakraborty et al., 2020) provide ever-improving means to drive the discovery processes toward good understanding of axon structure, which will eventually have important implications for work on neuronal pathologies.

Fig. S1 shows the homology of Drosophila Futsch to human MAP1B and Futsch-like genes in other invertebrates. Fig. S2 shows the composition of NFs. Table S1 contains ultrastructural reports of AISs in different neuron types. Table S2 contains axonal MT and NF numbers as reported in the literature.

The author is most grateful to Christophe Leterrier and Anthony Brown for constructive feedback and most valuable comments and advice; Yanhe Li for sharing the genomic sequence of crayfish; and Llilians Calvo-González for extracting fragments homologous to Futsch; Paul Conduit, Matthew Ronshaugen, and Andre Voelzmann for feedback; Samuel Shields for kind support with literature searches; and the Manchester University Library for their outstanding services.

Work underpinning this article was made possible through support by the Biotechnology and Biological Sciences Research Council to A. Prokop (grants BB/I002448/1, BB/P020151/1, BB/L000717/1, and BB/M007553/1).

The author declares no competing financial interests.