1.1. The structure of the axoneme comes into view

Our understanding of axonemal structure is intertwined with developments in electron microscopy. Early transmission electron-microscopy (TEM) studies, using thin-sectioned, resin-embedded samples, described the canonical 9 + 2 microtubule architecture of the axoneme of motile cilia, with nine doublet microtubules (DMTs) encircling a central apparatus (CA) of two singlet microtubules (Taira & Shibasaki, 1978; Afzelius, 1955; Fawcett & Porter, 1954; Fig. 1a). Similar experiments on nonmotile (primary) cilia revealed a 9 + 0 architecture comprising DMTs without the CA (Barnes, 1961). These and further studies provided the first views of the DMT, revealing an A tubule with 13 protofilaments connected to a B tubule with ten protofilaments. For motile cilia, they also revealed periodic complexes projecting from the microtubule surfaces (Fig. 1b). Each axonemal DMT was shown to be composed of 96 nm repeating units that feature two rows of dynein motors – outer and inner dynein arms – that power the ciliary beat. Additional regulatory complexes such as radial spokes (RSs) and the nexin–dynein regulatory complex (N-DRC) form connections with adjacent dyneins to adjust the waveform dynamics in response to mechanical cues (Witman et al., 1978; Bozkurt & Woolley, 1993).

Figure 1 The use of electron microscopy to visualize ciliary axonemes. (a) Resin-embedded, thin-section transmission electron microscopy (TEM) revealed the 9 + 2 architecture of many axonemes, in which nine doublet microtubules (DMT) encircle a central apparatus (CA) of two singlet microtubules. (b) Thin-section TEM also revealed periodic projections from both the CA and DMT. These images were obtained from https://commons.wikimedia.org. (c) Cryo-ET and subtomogram averaging of axonemes revealed the architecture of the 96 nm repeat unit of the DMT, providing the location and shape of many axonemal complexes including outer dynein arms (ODAs), inner dynein arms (IDAs), radial spokes (RS), the nexin–dynein regulatory complex (N-DRC) and the tether/tetherhead (T/TH) complex. EMDB entry EMD-2115 is shown colored by complex (Bui et al., 2012). (d) Single-particle analysis cryo-EM of splayed axonemes from C. reinhardtii allowed an atomic model of the 96 nm repeat to be built (PDB entry 8glv; Walton et al., 2023).

Later work eliminated the need for chemical fixation by imaging cryopreserved cilia and ciliated cells (Nicastro et al., 2005, 2006; Sui & Downing, 2006). Because of the size and the intricate architecture of the axoneme, the first cryoelectron microscopy (cryo-EM) studies of Chlamydomonas reinhardtii and sea urchin sperm axonemes were studied by tomography (cryo-ET), in which the samples were imaged from multiple angles to create a tilt series that could be aligned and interpolated to create a three-dimensional tomogram. Subtomogram averaging (STA) exploited the periodicity of the DMT to enhance the resolution, improve signal-to-noise ratios and mitigate anisotropy. These refined averages were the first to distinguish the orientation of the dyneins and their connections to other axonemal complexes (Nicastro et al., 2005, 2006; Bui et al., 2012; Fig. 1c), providing valuable information about the regulatory mechanisms that coordinate dynein activity. In addition to these large features, tomographic reconstructions also identified densities periodically associated with the inside walls of the DMT now known as `microtubule inner proteins' or `MIPs' that repeat with a periodicity of up to 48 nm (Sui & Downing, 2006; Nicastro et al., 2006). These were initially speculated to reinforce the protofilament network and dock axonemal complexes, a prescient suggestion that was later substantiated by functional and structural studies (Nicastro et al., 2006; Owa et al., 2019; Ichikawa et al., 2017, 2019).

These pioneering cryo-ET studies significantly improved our understanding of axonemal ultrastructure. However, it was only upon the introduction of cryo-EM with single-particle analysis (SPA) that subnanometre reconstructions became possible. SPA removed the need to tilt the sample during data acquisition, which led to higher throughput, less electron damage, improved signal-to-noise ratios and, ultimately, higher resolution. Applied to the DMT, advances in high-resolution imaging, sample preparation and particle-alignment workflows culminated in the reconstruction of the 48 nm repeat of the C. reinhardtii DMT at ∼3.5 Å resolution (Ma et al., 2019) and the generation of an atomic model (Fig. 1d). Once this was achieved, the application of SPA cryo-EM to DMT complexes from various species and cell types resulted in a boon of reconstructions that have proven to be key in establishing the identities, positions, repeat lengths, structures and interactions of DMT-associated proteins (Gui, Farley et al., 2021; Walton et al., 2023; Kubo et al., 2023; Leung et al., 2023; Zhou et al., 2023; Gui, Croft et al., 2022). Information from these structures has provided invaluable insights into the function of axonemal proteins, the evolution of the axoneme and the molecular underpinnings of ciliopathies.

1.2. Reconstruction of axonemal microtubules

There are several hurdles that make determination of the structure of axonemal microtubules by SPA cryo-EM difficult. Imaging the axoneme directly produces projection images with overlapping DMT and CA microtubules, complicating particle picking and alignment (Cheng et al., 2015). Methods for deconstructing the axoneme into its constitutive filaments have proven to be successful in separating axonemal microtubules, but typically result in the degradation or loss of native interactions. To overcome the challenge of sample preservation, axonemes can be gently splayed into constitutive filaments suitable for cryo-EM imaging while retaining the majority of axonemal proteins (Walton et al., 2021, 2023).

In addition to the difficulties introduced by sample preparation for cryo-EM imaging, particle alignment of DMTs is not straightforward, mainly due to their 96 nm repeat units. During particle picking, the register of these units along the microtubule filament is unclear, so any given point along a DMT filament will be at an indeterminate position within the 96 nm repeat. Therefore, to accurately align particles with the correct periodicity, the local register of the particles must be identified through a series of 3D classifications starting from 8 nm picked particles. As a result, particles can be aligned with periodicities of 24, 48 or 96 nm, which are sufficient to resolve the majority of DMT-associated proteins.

The last challenge in determining the structure of DMTs is the construction of atomic models. This process involves identifying axonemal proteins, generating initial models, integrating them within the larger complex and refining them against the cryo-EM density. Because the axoneme is a complex comprising several hundred different proteins, many of whose structures have not been determined, each step in this process is demanding. However, the challenge of model building has waned with improvements in automated, AI-driven model-building methods.

This review will summarize the advances that made atomic models of axonemal microtubules possible. We will describe sample preparation for SPA imaging, review the image-processing procedure and discuss methods for protein identification and model building for cryo-EM density. Finally, we will evaluate the exciting new developments that make cryo-ET poised to re-emerge as the premier method for axonemal structure determination.

2.1. Sample preparation for SPA imaging

Early studies of microtubule structure by cryo-EM relied upon in vitro reconstitution (Nogales, 2015); however, this approach is unsuitable for DMTs or the CA due to the large number of microtubule-associated factors and their complex patterns of localization. Instead, studies of axonemal structure have traditionally used native isolation techniques, leveraging the exceptional stability of ciliary microtubules. Motile cilia can be removed from cells by a combination of dibucaine treatment, millimolar levels of Ca²⁺ or mechanical force, and then crudely separated from cell debris by centrifugation over sucrose cushions (Craige et al., 2013; Beneke et al., 2020). Mild detergent treatment followed by pelleting yields bare axonemes, which must be further broken apart before imaging by SPA to avoid densely overlapping microtubules. The earliest studies of the Tetrahymena DMT by cryo-EM broke the axoneme apart using a combination of ATP and NaCl washes (Ichikawa et al., 2017; Maheshwari et al., 2015). ATP activates the axonemal dyneins to help slide apart the microtubules of the axoneme, while the high-salt washes remove the axonemal dyneins and break the electrostatic interactions that hold neighboring DMTs together. In addition, radial spokes were removed using dialysis of the pelleted DMTs after salt extraction and the DMTs were broken into fragments by sonication.

An alternative to NaCl is to use mild protease digestion to sever the linkages between DMTs, followed by the addition of ATP to telescope apart the axoneme through DMT sliding. This process, called sliding disintegration, has historical precedent in the seminal work of Summers & Gibbons (1971), which corroborated observations by Satir that DMTs move relative to each other during ciliary beating (Satir, 1967). Throughout the decades, a variety of proteases, including subtilisin A, which cleaves the tubulin C-terminal tail (Sackett et al., 1985), and reaction conditions have been used to digest the axoneme for biochemical and structural studies. To minimize the damage caused by digestion, sliding disintegration is often performed on ice in the presence of protease inhibitors. This approach yielded the first sub-4 Å resolution structure of the DMT 48 nm repeat (Ma et al., 2019) and, soon after, the radial spoke and CA complexes (Gui, Ma et al., 2021; Gui, Wang et al., 2022).

2.2. Cryo-EM image processing

To reconstruct the DMT, a specialized image-processing pipeline is required due to the periodic nature of axonemal complexes bound to the tubulin surface. This processing differs from the microtubule-reconstruction methods developed for cytosolic microtubules (Zhang & Nogales, 2015; Cook et al., 2020), which are primarily concerned with determining the number of protofilaments (as cytosolic and especially in vitro reconstituted microtubules can have variable numbers of protofilaments) and the location of the microtubule seam (a break in the helical symmetry of the filament where heterotypic α–β and β–α lateral contacts occur) (Chrétien & Wade, 1991; Kikkawa et al., 1994; Ray et al., 1993; Kellogg et al., 2017). Neither of these concerns are particularly relevant to DMTs, which have a set number of protofilaments distributed between A and B tubules, and large projections that create features that preclude the need to specifically locate the seam. In this section, we will outline the image-processing workflow used to reconstruct the axonemal DMT (Fig. 2).

Figure 2 The cryo-EM processing scheme used to reconstruct doublet microtubules (DMTs). (a) Electron micrographs showing splayed axonemes at medium magnification (left) and DMTs (each marked with an arrowhead) at high magnification (right). The DMTs are picked (illustrated with a yellow line for a single DMT) prior to particle extraction. (b) Schematic showing the steps taken to reconstruct DMTs at progressively higher periodicities, starting with particles extracted every 8 nm along the longitudinal axis of the microtubule. Each step relies on mask-focused classification with masks (illustrated as yellow cylinders) covering features known to have a higher periodicity than the map to which they are applied. For example, 16 nm periodicity can be obtained by classifying on the position of FAP52 (a microtubule inner protein with 16 nm periodicity) in the 8 nm reconstruction. (c) The quality of the initial alignment can be monitored by inspecting the inner junction: a repeating pattern of FAP20 and PACRG. These proteins should be clearly distinguishable from one another in well aligned 8 nm reconstructions.

3.1. Protein identification pre-AlphaFold

The first ∼3.5 Å resolution structure of a DMT presented a formidable model-building challenge (Ma et al., 2019). Although the tubulin component could easily be modeled by substituting the sequence from previously solved cytosolic microtubule structures and manually refitting the model to the map, the dozens of MIPs had to be identified directly from the cryo-EM density. For globular MIP densities, the MOLREP–BALBES pipeline, a density-based fold-recognition workflow (Brown et al., 2015; Vagin & Teplyakov, 2010), was used (Fig. 3a, left). In this workflow, MIP densities were extracted from the composite map and each entry in the BALBES database (Long et al., 2008), a database of unique protein folds derived from the Protein Data Bank (PDB), was systematically fitted to the extracted densities using MOLREP (Vagin & Teplyakov, 2010). The solutions were then ranked and the top hits manually inspected to identify the best-fitting domain. Cross-referencing the recognized fold with mass-spectrometric analysis of the cryo-EM sample typically identified a MIP with a similar domain, while side-chain analysis differentiated the correct solution. This semi-automated approach was successful in identifying FAP67, FAP161, RIB72, FAP363, FAP182 and multiple EF-hand motifs (Ma et al., 2019). Once the proteins have been identified in one species, homology-modeling programs such as SWISS-MODEL and I-TASSER (Schwede et al., 2003; Zhang, 2008) can be used to build starting models for DMTs from other species.

Figure 3 Strategies for building atomic models. (a) General scheme for interpreting cryo-EM maps of unknown identity before 2021. The protein fold responsible for a region of globular density could be determined by evaluating the fit of a library of protein folds (the BALBES database) curated from the PDB to the density (Brown et al., 2015; Vagin & Teplyakov, 2010). Map regions with resolved side chains could be built manually, with side-chain predictions used to generate query sequences for BLAST searches against proteomes. Alternatively, secondary structure could be extracted from the built model and compared with precomputed profiles, or the tertiary structure compared with PDB entries using DALI (Holm, 2022) or PDBeFold (Krissinel & Henrick, 2004) to identify likely folds. (b) Current best approaches for interpreting cryo-EM maps. The protein fold responsible for a region of globular density can be determined by evaluating the fit to the density of AlphaFold2-created model libraries of entire proteomes (Jumper et al., 2021). Map regions with resolved side chains can be built automatically using AI-based tools such as ModelAngelo (Jamali et al., 2024) or DeepTracer (Pfab et al., 2021; Chang et al., 2022). Sequence assignments can be extracted and used to create profile hidden Markov models (HMMs) for proteome searches. Alternatively, the three-dimensional models can be compared with millions of predicted structures using Foldseek (van Kempen et al., 2023) to identify the exact protein or protein fold.

Nonglobular and remaining MIP densities were built de novo (Fig. 3a, right), starting with polyalanine traces built manually in graphical programs such as Coot (Emsley et al., 2010). These initial traces were then used as templates to assign side chains or secondary structure: MIPs could be identified by comparing the assigned sequence against proteomes (either of the whole organism or a subset of proteins obtained from mass-spectrometric analysis of the sample) or comparing the secondary-structure profile with predictions. The traced models could also be used as an input for search programs such as PDBeFold (Krissinel & Henrick, 2004) and DALI (Holm, 2022) to identify structurally similar proteins pre-existing in the PDB. As in the example of the MOLREP–BALBES pipeline, knowledge of the protein fold can be used to identify proteins with similar domains in the mass-spectrometric results, simplifying the search space. The recent development of Foldseek means that this approach can now be used to search millions of predicted structures (van Kempen et al., 2024).

3.2. Developments in protein identification and model building

More recently, advances in protein-structure prediction and automated model building have upended these past workflows to allow the rapid identification of unknown protein densities (Fig. 3b). Arguably the most transformative advance has been the development of AlphaFold2 (Jumper et al., 2021) and similar neural network-based methods for structure prediction. The ability to rapidly predict protein structures directly from primary sequences now makes it feasible to create or simply download structural libraries of nearly every protein in the proteome of an organism. This expansion of confident protein predictions can be leveraged to identify unknown density in cryo-EM maps. For example, the BALBES database can be substituted with an AlphaFold2 library in domain-recognition workflows, providing a way to perform domain-based recognition with entire proteomes (Fig. 3b, left). This approach, using COLORES (Wriggers et al., 1999) as the search program, was recently used to identify CCDC105 and other MIPs in the cryo-ET density of the mouse sperm DMT (Chen, Shiozaki et al., 2023; Fig. 4). Importantly, the success of fold recognition is often more dependent on the quality of the starting models than having high-resolution features. AlphaFold2 has therefore opened the door to identifying axonemal proteins within lower resolution maps, although without side-chain-level details the risk of misidentification remains, especially among paralogous proteins.

Figure 4 Subtomogram average of the mouse sperm doublet microtubule (DMT) and identification of CCDC105. (a) Cross-sectional view of the 48 nm repeat of the mouse sperm DMT determined using in situ cryo-ET and subtomogram averaging (EMDB entry EMD-41431; Chen, Shiozaki et al., 2023). The overall resolution is estimated to be 7.7 Å. CCDC105, a microtubule inner protein of the B tubule, is shown in orange. (b) Longitudinal view of density corresponding to CCDC105 extracted from the DMT map. (c) An atomic model of CCDC105 showing its 16 nm periodicity. CCDC105 was identified using an unbiased search with an AlphaFold2 library containing 21 615 predicted models of the mouse proteome. CCDC105 has been confirmed at this location in other sperm DMTs using single-particle analysis cryo-EM (Leung et al., 2023; Zhou et al., 2023). (d) AlphaFold2 model of CCDC105.

Equally as impactful has been the emergence of de novo model-building software such as DeepTracer (Pfab et al., 2021; Chang et al., 2022) and ModelAngelo (Jamali et al., 2024) that exploit deep-learning methods and can build into any cryo-EM density with resolved side chains, regardless of sequence information (Fig. 3b, right). One of the advantages of these approaches, as well as model-building aids such as findMy­Sequence (Chojnowski et al., 2022), is that they calculate probabilities for all 20 amino acids for each built residue of every chain. These probability profiles provide better queries for protein identification than single solutions (Mistry et al., 2013). The power of this approach was illustrated by the identification of four radial spoke proteins (now referred to as RSP24, RSP25, RSP26 and RSP27) and two CA proteins (FAP92 and FAP374) that were unable to be assigned using traditional approaches (Jamali et al., 2024). The development of AlphaFold2 and automated model-building software have not only impacted the field of cilia biology but also the structural biology of large protein assemblies in general.

3.3. Model refinement of axonemal complexes

Refining atomic models of axonemal complexes remains a substantial challenge because of their size and the fact that the composite maps exhibit a large range of local resolutions. For the 96 nm repeat of the DMT, its large size prevents refinement of the complex as a single model even after reducing the map dimensions to the smallest cuboid capable of containing the full length of the repeat. Because of these issues, models of 96 nm axonemal assemblies currently need be split into local regions and refined into the cryo-EM density as separate subregions. To allow the refined models to be recombined into a full assembly, these subregions must overlap to avoid the introduction of unrefined interfaces.

4.1. Using cryo-ET to study axonemes at high resolution

By the mid-2010s, SPA cryo-EM could yield resolutions of ∼3 Å for protein complexes, yet cryo-ET and subtomogram averaging was limited to subnanometre resolution only for large, highly symmetric complexes in thin ice (Kühlbrandt, 2014). The additional inelastic scattering generated by thicker ice (>100 nm) and the electron damage caused by repeated imaging during tilt-series acquisition precluded high-resolution reconstructions for most cryo-ET samples. A leap towards subnanometre-resolution reconstructions of the axoneme was achieved in 2019 by combining the strengths of cryo-ET with SPA cryo-EM through a process called Tomography-Guided 3D Reconstruction of Subcellular Structures or TYGRESS (Song et al., 2020). TYGRESS involves acquiring two data sets per target: firstly a single image at 0° tilt with a high dose and secondly a traditional tilt series captured at a lower dose. Following collection, the tilt series is used to create sub­tomogram averages, from which 3D coordinates and alignment angles are extracted. The 3D coordinates are subsequently used to pick particles from the 0° tilt, high-dose image and obtain alignment angles to initialize refinement with SPA methods. Therefore, using this method, the 3D position and initial alignments of each particle are derived from the tilt series, while the higher signal-to-noise ratio 0° initial image is used for high-resolution reconstruction. This general method is particularly well suited to determining structures from highly crowded micrographs. Applied to the T. thermophila axoneme, TYGRESS was able to reconstruct the DMT to an average resolution of 12 Å, a significant improvement at the time that clearly resolved individual subcomplexes and some large MIPs (Song et al., 2020). Although impressive, these reconstructions did not allow accurate atomic model building.

Subsequent improvements in cryo-ET methodology have produced several subnanometre reconstructions of sperm DMTs, including ∼5 Å resolution reconstructions of mouse sperm DMTs (Chen, Shiozaki et al., 2023; Chen, Greenan et al., 2023; Tai et al., 2023; Fig. 4a). This work has opened the possibility of studying ciliary axonemes by cryo-ET at high resolution. One major advance was the introduction of focused ion-beam (FIB) milling of cryopreserved whole sperm cells to create electron-penetrative lamella. This results in intact axonemes in relatively thin ice with improved signal-to-noise ratios. Axonemes prepared in this way are also typically less `flattened' than when prepared by freezing isolated cilia directly onto cryo-EM grids (Tai et al., 2023), helping to preserve the geometry of the axoneme and limit damage to the DMTs. Additionally, cryo-ET image-processing workflows have improved by drawing on principles from SPA studies. For studies of DMTs that have resulted in reconstructions below 10 Å, alignments of subtomograms were performed in RELION-3 or RELION-4, which use a Bayesian statistical approach to align particles based on maximizing the likelihood given the observed data and prior information (Zivanov et al., 2018, 2022). Recent versions of RELION include cryo-ET-specific tools, such as CtfRefineTomo and TomoFrameAlign, that promise to improve reconstructions further. Excitingly, this resolution breakthrough promises to improve our understanding of axonemal structures in their native environment.

4.2. Cryo-ET and axoneme asymmetry

With the dawn of high-resolution cryo-ET imaging of intact cilia, native features such axoneme asymmetry may be observed. Asymmetry underlies many important features of axonemes in different species, such as the DMT–fibrous sheath connections in mammalian sperm, the DMT–para­flagellar rod connections in trypanosomatid-family parasites, and several noted radial asymmetries in C. reinhardtii (Dutcher, 2020; Bastin et al., 1998; Baccetti & Afzelius, 1976). Unfortunately, one of the limitations of SPA cryo-EM is that by deconstructing the axoneme into its constituent filaments, the 9 + 2 organization is destroyed and asymmetrical features of distinct DMT microtubules are averaged and lost. Cryo-ET of native axonemes avoids this issue. One recent example is the cryo-ET study of FIB-milled mouse sperm axonemes, in which separate reconstructions of each DMT revealed that only some have radial spokes associated with a large barrel that may have a role in regulating the sperm-specific asymmetric waveform (Chen, Greenan et al., 2023). Although axoneme asymmetries have been observed for decades, their molecular basis and function are only now coming into focus through recent improvements in resolution and map quality.

4.3. Uncovering transient structures on the axoneme

Cryo-ET has also been central to determining structures of the intraflagellar transport (IFT) A and B cargo-trafficking complexes, which transiently associate with the axoneme. These complexes form long oligomers, termed trains, that assemble at the base of the cilium and traffic cargo to the tip in a kinesin-dependent translocation on DMTs called antero­grade trafficking. Upon reaching the tip, the train reassembles and returns to the base in a dynein-dependent process termed retrograde transport. Tomographic analysis has now observed IFT trains in C. reinhardtii assembling at the base of the cilium (van den Hoek et al., 2022) and during active anterograde transport (Lacey et al., 2023; Jordan et al., 2018). Subtomogram averaging applied to the anterograde train achieved a nearly subnanometre resolution, allowing model building (Lacey et al., 2023) that was supported by contemporary single-particle methods (Meleppattu et al., 2022; Jiang et al., 2023; Ma et al., 2023). This work opens the door to capturing the structures of active processes taking place on the axoneme. Future structures of the retrograde IFT train and trains associated with specific cargos have the potential to offer enormous insight into the process of protein trafficking in cilia.

4.4. Cryo-ET and primary cilia

Cryo-ET, and the wider techniques of volume electron microscopy (vEM), also provide suitable methods to image primary cilia. Although it is a convention to describe primary cilia as having a 9 + 0 architecture, tomography on primary cilia isolated from MCDK-II cells (Kiesel et al., 2020) and serial section-imaging studies of IMCD3 kidney cells (Sun et al., 2019) and C. elegans primary cilia (Doroquez et al., 2014) have demonstrated that the structures of these axonemes are unconventional, with large radial differences in their organization and DMTs that rapidly taper into singlet microtubules. Future subtomogram averaging of the microtubules of primary cilia should provide new information about their microtubule-associated proteins, about which very little is currently known. Because cells typically have only one primary cilium, fluorescent markers may be useful to help focus imaging on intact FIB-milled primary cilia.