Imaging of retina cellular and subcellular structures using ptychographic hard X-ray tomography

The sensory input for vision is located in the retina, which is part of the central nervous system. The retina contains all known photoreceptors required for vision and the majority of non-visual photoreceptors, as well as a complex network of neurons and specialized cells for signal transmission and computing. It is one of the most studied and most imaged neural tissues. These studies contribute to major advancements in our understanding of intricate neuronal networks (Wässle, 2004), for example on the visual perception of motion (Behnia et al., 2014; Wernet and Desplan, 2014) and direction selectivity (Yonehara et al., 2016). A variety of imaging techniques have been key to understanding the process of vision, and its pathological counterpart, at different tissue organization levels, intercellular and subcellular, which are orderly organized in consecutive cells layers. The retina, site of the cell circuit photoreceptor-bipolar-ganglion cells, the pathway for signal transfer to the brain, is classically subdivided into seven sublayers called the outer segment (OS), inner segment (IS), outer nuclear layer (ONL), outer plexiform layer (OPL), inner nuclear layer (INL), inner plexiform layer (IPL) and ganglion cell layer (GCL) (Wässle, 2004) (Fig. 1; Fig. S1). Electron microscopy (EM) has earlier enabled an overall 3D impression of the neural cells interconnecting along a layer, using freeze-fracture scanning EM (Wässle, 2004). However, EM is most efficient for analyzing subcellular structures at highest resolution, such as the outer segment disks of photoreceptors generated at the cilium (Anderson and Fisher, 1975) and the sophisticated multi-synapses complexes in triad or tetrad organization at the photoreceptor ends (Haverkamp et al., 2000). Most neurobiology studies with 3D imaging use targeted fluorescent labeling of subcellular domains, lacking overall information of the unstained cell population. The optimization of automated EM for 3D imaging, using serial block-face scanning EM, has allowed the detailed analysis of the overall ultrastructure of some retina layers in healthy specimens or those presenting degenerative phenotypes (Woodell et al., 2013; Pollreisz et al., 2018, 2020; Annamalai et al., 2020). A challenging 3D mapping of the neuronal network of the mouse IPL using serial block-face EM (Helmstaedter et al., 2013) allowed the imaging of 950 neurons in a cubic volume of 100 µm side length and established a procedure for the classification, localization and connectivity of the neurons. However, this method involves sectioning of the sample, thereby destroying the sample and preventing further investigations.

The need for non-destructive imaging of volumes of tissue in 3D, which allows for identification of regions of interest for further study, using, for example, correlative microscopy or proteomic/genomic analyses, drove the development of new imaging techniques (Karreman et al., 2017). Such a method is also highly desirable in imaging retina tissue, for analyzing the neuron circuitry and the neural fine structure in healthy and pathological tissues. X-ray imaging of biological structures is highly complementary in terms of sample size and spatial resolution to other imaging techniques such as EM and light microscopy. Owing to the excellent penetration behavior of X-rays into matter, thicker samples can be imaged, albeit with limited resolution (Du and Jacobsen, 2018). Soft X-ray full-field microscopy can be used to image entire cells of up to ∼10 µm with a resolution allowing the visualization of subcellular structures in detail (Foglia et al., 2019; Guo and Larabell, 2019). Hard X-ray beams, with energies above 2 keV, are needed for imaging thicker biological samples up to several hundreds of microns (Krenkel et al., 2015). Phase-contrast modalities, such as propagation-based methods (Kosior et al., 2012; Krenkel et al., 2015; Töpperwien et al., 2018; Kuan et al., 2020), need to be used due to the poor absorption contrast of organic matter in water at these energies. A method using hard X-ray ptychographic nanotomography, also called ptychographic hard X-ray computed tomography (PXCT) (Dierolf et al., 2010), has recently been used to image a 10 μm-thick piece of a computer microprocessor at a resolution down to 15 nm in 3D (Holler et al., 2017a). PXCT has also been successfully used for biological samples, at cryogenic temperatures required at the typical energies used, first with the imaging of resin-embedded, stained plant tissue (Guizar-Sicairos et al., 2011) and for imaging frozen cells, revealing subcellular structures in 3D (Diaz et al., 2015). The development of specific instrumentation with a cryogenic sample holder (Chen et al., 2014; Holler et al., 2018) and high-accuracy sample positioning has provided tomograms of chemically fixed frozen-hydrated cells (Deng et al., 2018) and chemically fixed frozen-hydrated brain tissue of ∼20×70×70 µm³ with a resolution of ∼120 nm (Shahmoradian et al., 2017; Tran et al., 2020).

Here, we present 3D imaging with PXCT of neuronal connections in the retina outer layer from wild-type mouse and VPP transgenic mouse (Naash et al., 1993), a model for retinitis pigmentosa (RP) characterized by progressive photoreceptor degeneration, with the aim of establishing a tool in neurobiology for retina degeneration analysis. A comparison of the imaging using PXCT and focused ion beam (FIB)/scanning electron microscope (SEM) tomography from the same region of the retina (Fig. S2) shows the excellent complementarity of both techniques.

A retina from a wild-type mouse or a mouse model of RP was aldehyde fixed, osmium stained and embedded in epoxy resin (Fig. 1A, bottom), as described for electron transmission microscopy in Heynen et al. (2013). A cylindrical sample of ∼25 µm diameter and 50 µm height was milled using FIB/SEM, in the OPL (Fig. 1A, top left) containing the plexus of terminal neuronal connections formed at the interface of the outer (photoreceptor cells) and inner (bipolar cells) retina layers. The region of interest for milling could be easily identified after SEM imaging of the retina surface, owing to the rough surface of the neighboring ONL (Fig. 1A, top right) made of photoreceptor nuclei. The FIB/SEM-milled sample was placed in the center of a gold-coated pin (Holler et al., 2017b) for X-ray imaging (Fig. 1B). The quality of the sample was verified by SEM after removing the protective carbon top layer (Fig. 1B, inset). For an adaptation of this technique to frozen-hydrated samples, see Shahmoradian et al. (2017).

The retina sample from the wild-type mouse described in Fig. 1, milled out of the OPL and slightly overlapping the ONL in order to have some rod nuclei, was measured over a height of 20 µm. The reconstructed tomogram shows the overall organization of the photoreceptor cells plexus in 3D at an isotropic half-period resolution of ∼200 nm (Fig. 2A; Movie 1). The reconstructed tomogram contains ∼60 photoreceptor cell nuclei.

Transversal orthoslice views across the layers (Fig. 2B,C) of the reconstructed tomogram show the nuclei of the ONL with the characteristic high-density chromatin (ch), the photoreceptors synapses (sy) and axonal fibers (ax), which can be traced efficiently using a 3D viewer and analysis software. Fig. 2D shows a 3D segmentation of a nucleus (magenta) and the path of the photoreceptor cells axonal fibers (green) with a diameter of ∼500 nm, crossing the ONL and reaching the OPL. Fig. 2E–G show a zoom on several synaptic ends, with the typical large mitochondria (m) close to the active high electron-dense zone (labeled ‘az’ for ‘active zone’). A longitudinal cut along the OPL of the retina reveals the synapses (Fig. 2H) populated by high-electron density regions (Fig. 2H, inset) being most likely synaptic clefts (az). This view also demonstrates the circularity of the synapses (Fig. 2H).

The OPL of the VPP mouse, a model for autosomal-dominant RP (Ding and Naash, 2006; Naash et al., 1993) was investigated (Fig. 3B,D) in order to compare the tissue with the structure of the wild-type mouse (Fig. 3A,C) and assess whether the method is suitable for analyzing specimen from retina with degeneration. Light microscopy images (Fig. 3A,B) show in the RP mouse retina the well-known loss of photoreceptors (Naash et al., 1996; Samardzija et al., 2006b) (Fig. S4), resulting in the deterioration of the rod outer segments (OS) and progressive thinning of the ONL (Fig. 3B). At this developmental stage of the disease, the OPL does not seem to be dramatically affected when observed using a classical low-resolution histology method (Fig. 3B). However, a closer look at the RP mouse OPL using PXCT reveals very pronounced degeneration processes compared to the wild-type OPL at the same age, and the layer no longer displays the rod synaptic ends (sy), called spherules, and axonal fibers (ax) (Fig. 3D) (Movies 1 and 2, Fig. S4). Widespread fibrosis can be seen in the 3D volume of the OPL, as well as bipolar cell bodies, vascularization and synapse remnants (Movie 2).

Analysis of the PXCT tomograms from the wild-type retina revealed many synapses (Fig. 4A, sy), with the strongest electron density at the synaptic ribbon (sr), dense assembly of proteins and membranes of several hundreds of nanometers (Migdale et al., 2003). The synaptic ribbons were actually the smallest structures that we could observe. Besides the usual subcellular structures of photoreceptor cells, a particular feature displayed was the strong electron density in the bodies of a few cells located at the ONL/OPL interface. They harbor a synapse directly in the soma (ssy) [Fig. 4C; Fig. S3, Movies 1 (overall tomogram, green arrow) and 3 (subvolume of the tomogram, single cell)], and are easily recognized by the strong electron-dense ribbon and hilus (h). The hilus (Migdale et al., 2003) or hilum is a complex at the plasma membrane region of the synaptic end, forming a common opening fold for the neurites to connect to the synapse.

As PXCT operates without sectioning, the integrity of the sample was preserved, and further analysis like imaging at higher resolution with another technique was possible (Fig. S2). Although EM usually requires multiple staining compared to the single staining used here, FIB/SEM tomography at the region of interest displayed enough contrast to visualize the photoreceptor nuclei, axonal fibers and synaptic ends, as well as the synaptic ribbon (Fig. 4B; Movie 4). It is of great advantage to have a fixation and staining protocol compatible for both PXCT and FIB/SEM imaging. In some cells, particularly at the ONL/OPL interface, we could also identify ribbon synapses in the soma, like we observed with PXCT. A somatic synapse was selected (Fig. 4D; Movie 5) and segmented (Fig. S5). Ribbon synapses have already been demonstrated in the soma of cochlear afferent mechanosensory hair cells (Wichmann and Moser, 2015) connecting to the auditory nerve fibers (Goutman et al., 2015), as well as in rod cells, in which they seem to be increased in Down syndrome (Li et al., 2015). The FIB/SEM tomogram shows that the synaptic ribbon of the cell body is also associated with a large mitochondrion (Fig. 4D, m), a classical feature of rod synaptic ends, and flanked by vesicles distributed in proximity to the presynaptic membrane. A bunch of dendritic vesicles of ∼30 nm trapped at a height between the ribbon and the hilus suggests functionality of this synapse (see yellow arrows in Fig. S5F).

Cryo-PXCT is emerging as a powerful 3D scanning technique in biology, and allows imaging from single cells (Diaz et al., 2015) to dozens of cells while maintaining the integrity of the sample (Shahmoradian et al., 2017). Here, imaging a resin-embedded retina sample (25 µm diameter and 20 µm height) using PXCT, we achieve a half-period spatial resolution of ∼200 nm, limited mostly by the small difference in density contrast between organelles and water with the available coherent X-ray flux. This resolution was sufficient to see large organelles and other intracellular structures like synaptic ribbons with a length of a few hundred nanometers. With its good depth of field (Table 1), PXCT allowed us to image ∼60 neuron cell nuclei, interlaced with axonal fibers traced with their synaptic ends until the plexiform layer.

The retina of a transgenic mouse, a model of RP, was analyzed by PXCT in the same region of interest and compared with wild-type retina. The tomograms obtained in this study demonstrate that PXCT is a highly suitable method for comparative studies on neurodegenerative disease evolution in tissues. Comparing retina of wild type with retina of RP, a clear degeneration of the tissue is observed in 3D, with a drastic reduction in synaptic ends of photoreceptor cells. This result paves the way for our further analysis of the evolution of retinal degeneration and remodeling in transgenic mouse models of retinopathy.

The PXCT method presented here is an excellent bridge between a low-resolution method that non-destructively scans large tissue volumes, i.e. in vivo hard X-ray microtomographic imaging at micrometer isotropic resolution (Walker et al., 2014), and the high-resolution imaging of thin samples, using SEM. PXCT does not require serial sectioning, thus allowing subsequent usage (Karreman et al., 2017), like the FIB/SEM tomography performed in this study. Therefore, correlative applications like high-resolution imaging or biochemical methods, such as proteomics, direct in situ mass spectrometry analysis and others, will be used in the future for tissue biology of complex structures in physiology and pathology disciplines, after PXCT scanning.

Conventional EM and PXCT are complementary in the resolution and accessibility of specimen volume (Table 1). PXCT achieves lower resolution but from larger volumes of biological tissue without sectioning. In addition, the methods also provide complementarity in the obtained information. When comparing PXCT with electron tomography datasets, we observe, on the very same sample and similar organelles, differences. Although PXCT and FIB/SEM do display maximum electron density signals on the same features, some other signals of lower intensity did not pinpoint exactly the same features with each method, i.e. at the synaptic ends. This property is most probably due to differences between the two methods with respect to the interaction of each beam with matter, and the type of detection. The local elemental contrast obtained with backscattered electrons in FIB/SEM differs from the electron-density contrast of PXCT (Diaz et al., 2015; Pfister et al., 2016). For example, at the synapses, the strongest observed intensities were, for both methods, located at the ribbon of the synapses, which are rich in proteins, and the hilus of the somatic ribbon synapse. However, although the membranes around cells and organelles were better resolved with the FIB/SEM method because osmium tetroxide interacts with lipids, PXCT highlighted particularly well and systematically electron-dense structures located e.g. at the junctions of synaptic arciform membranes downstream of the synaptic ribbon (Fig. S3E–G). The aforementioned property of hard X-ray tomography should be further exploited with higher resolution and contrast, and in a correlative way. It will likely allow pinpointing of supramolecular assemblies, also those barely present in a tissue, according to their density in samples without contrasting agent.

Photoreceptor cells are ∼100 µm long and usually possess a long axonal fiber traveling until the plexus part to reach the postsynaptic ends of the next neuron layer. Our first attempt to image a retina sample of 25 µm diameter and 20 µm length using PXCT allowed the analysis of a volume containing 60 nuclei and synaptic connections. We expect that, in the future, improvements in sample preparation (less handling, less staining, with fiducial and correlative options), in the PXCT method and algorithms (Li and Maiden, 2018; Odstrčil et al., 2019; Thibault et al., 2014; Tsai et al., 2016), and in the synchrotron X-ray beam properties (more coherent flux available in next-generation synchrotron sources) (Thibault et al., 2014), will enable imaging of volumes larger than 100×100×100 µm³ containing hundreds of cells at an isotropic spatial resolution of 50 nm.

C57Bl/6 wild-type and VPP mice were maintained as breeding colonies at the Laboratory Animal Services Center (LASC) of the University of Zurich in a 14 h: 10 h light–dark cycle with lights on at 06:00 and lights off at 20:00. Mice had access to food and water ad libitum. All procedures concerning animals were in accordance with the regulations of the Veterinary Authority of Zurich and with the statement of ‘The Association for Research in Vision and Ophthalmology’ for the use of animals in research.

Enucleated eye globes (Fig. S2) from wild-type mice or from VPP mice by 4 weeks of age (RP-model mice) (Naash et al., 1993) were fixed by 2.5% glutaraldehyde in 0.1 M cacodylate buffer pH 7.2 for 12 h at 4°C and trimmed, and the isolated retina was cut through the optic nerve into two parts, superior and inferior. Both retina parts were embedded in epoxy resin. The retina samples were stained for 60 min in 1% osmium tetroxide (postfixation) in 0.1 M cacodylate buffer at pH 7.2 at room temperature, stepwise dehydrated with increasing ethanol concentrations (15 min each for 30, 50, 70, 90 and 100%), followed by two supplementary 15 min incubations in 100% ethanol and embedded in Epon™ 812 resin (Samardzija et al., 2006a). Microtome sections (Leica Ultracut UCT) of 50 nm analyzed by widefield light microscopy (Axioplan Microscope, Zeiss, Jena, Germany) confirmed the quality of the staining, the preservation of the neuronal layers (Fig. S1) and the cell integrity (Fig. 1A, top left). The surface selected for further 3D imaging was coated with a 20 nm gold layer using a Balzers SCD050 sputter coater to make it conductive.

The block of resin-embedded retina originating from the superior half was mounted on the stage of a FIB/SEM NVision 40 electron microscope (Zeiss), and the region of interest was localized by SEM viewing mode using an SE2 (type 2 secondary electrons) detector and an energy of 2.00 kV (Fig. 1A, top right). The ONL is the easier layer to recognize because of the highly scattering nuclei. In order to respect the limits of volume imaging and time for data acquisition, a sample of ∼25 µm diameter and 50 µm height was milled (annular milling with 3 nA, then polishing with 700 pA) in the neighboring OPL (Fig. 1A, top left) or in the IPL (not shown) using the FIB of the NVision 40 microscope. The sample was prepared from a region close to the optical nerve (a few hundreds of microns away). The resulting cylindrical sample was transferred and welded with carbon, centered on a gold-coated pin for X-ray imaging (Holler et al., 2017a) (Fig. 1B) and stored at room temperature. The quality could be verified at higher resolution by SEM using an energy-selective backscatter (ESB) electron detector at an energy of 5.00 kV (Fig. 1B, inset).

Resin-embedded retina samples from wild-type and RP-model mice were measured using a ‘tOMography Nano crYo’ (OMNY) instrument (Holler et al., 2012, 2018) operating at the cSAXS beamline of the Swiss Light Source at the Paul Scherrer Institute in Villigen, Switzerland. The OMNY microscope (Holler et al., 2018) provides ideal conditions for 3D imaging with PXCT: ultra-high vacuum, cryogenic temperature control and, most importantly, accurate positioning of the sample controlled by laser interferometry (Holler and Raabe, 2015). The samples were mounted on a cryo-stage and were kept in a vacuum at a constant temperature of ∼90 K during the entire measurements. The purpose of using a cryo-stage is to keep the sample stable due to the possible radiation damage caused by X-rays (Gianoncelli et al., 2015; Garman and Weik, 2019) on the nanoscale. The measurement details of the wild-type specimen were reported in an earlier work (Holler et al., 2018).

Briefly, we used an energy of 6.2 keV and we coherently illuminated a Fresnel zone plate (FZP) made of Au (Gorelick et al., 2011) to define a coherent illumination onto the specimen with a size of ∼4 μm and a flux of ∼2.6×10⁸ photons/s. Details about the FZP and exact distances can be found in Holler et al. (2018). Ptychographic scans following the pattern of a Fermat spiral (Huang et al., 2014) were done with a field of view of 35×25 μm² (horizontal × vertical) and an average step size of ∼1.2 μm. At each scanning position, diffraction patterns of 0.1 s acquisition time were recorded with a Pilatus 2M detector (Henrich et al., 2009; Kraft et al., 2009) placed at 7.335 m downstream of the sample. Ptychographic scans were collected at 333 equally-spaced angular steps between 0 and 180 degrees, taking a total time of 16.5 h, including overhead time between acquisitions. We estimate that a dose of about 1×10⁷ Gy was absorbed by the sample during the entire measurement.

The RP-model mouse specimen was also measured at a photon energy of 6.2 keV, with the beam also defined by a FZP with 60 nm outer-most zone width. The diameter of the FZP was, in this case, 220 µm; thereby a coherent photon flux of 4.2×10⁸ photons/s was achieved. The sample was placed at 2.4 mm after the FZP, where the size of the beam was ∼8 µm. At this position, the specimen was scanned following the pattern of a Fermat spiral over an area of 60×20 µm² (horizontal×vertical) with an average step size of ∼2 µm. We used an Eiger 500 k detector (Dinapoli et al., 2011) placed 7.314 m downstream to record diffraction patterns at each scanning position with an acquisition time of 0.1 s. Scans were performed at 500 different angular orientations, obtained by rotating the specimen at different angles from 0 to 180° in equal angular intervals. The total acquisition time for this specimen was ∼9.7 h, imparting a dose of ∼6.1×10⁶ Gy on the specimen. In both samples, the dose was estimated from the incident X-ray fluence, as reported (Howells et al., 2009), assuming that the sample is mostly composed of resin. However, it is important to note that we used osmium as contrasting agent and that the absorbed dose is dependent on the distribution of the stain.

The diffraction patterns acquired for each ptychographic scan at each angular position were fed into iterative phase retrieval algorithms, more precisely a combination of the difference map (Thibault et al., 2009) and the maximum likelihood algorithms (Thibault and Guizar-Sicairos, 2012) was used. These reconstructions yielded 2D complex-valued images corresponding to each angular projection. The phase of these projections was extracted and corrected for zero and first-order phase terms, which are intrinsic degrees of freedom in ptychographic reconstructions (Guizar-Sicairos et al., 2011), and registered before tomographic reconstruction as described in Guizar-Sicairos et al. (2011) and Guizar-Sicairos et al. (2015). The size of the diffraction of patterns used for ptychographic reconstructions was different for each sample: in the case of the wild-type specimen, the diffraction patterns were cropped to 400×400 pixels of 172 µm pixel size; for the RP-model sample, 500×500 pixels of 75 µm pixel size were used. This determined the resulting pixel size in the reconstructed ptychographic projections, which is equal to the voxel size in the tomographic reconstructions, to be 21.9 nm and 39.0 nm for the wild-type and RP-model specimens, respectively. The spatial resolution of the tomograms was estimated using Fourier shell correlation (FSC). We then computed the FSC between these tomograms, which is a correlation curve between the two 3D datasets in Fourier domain. The resolution was estimated using as threshold the half-bit criterion (van Heel and Schatz, 2005). For this analysis, the two independent tomograms were cropped to a 3D region inside the sample containing all cell nuclei, synaptic connections and other structures beyond the synaptic ends, which were not well resolved. The average estimated half-period resolution within such a volume was ∼150 nm and 230 nm for the wild-type and RP-model samples, respectively. The difference in the two values stems from the lower contrast present in the features of the RP-model sample.

The EM images on the very same control sample confirm that no damage has occurred at this resolution on the ultrastructure of the retina tissue due to X-ray exposure. A high X-ray fluence on resin-embedded, stained biological tissues can additionally cause a deformation of the sample during the acquisition, which can affect the 3D resolution when combining all projections to reconstruct a tomogram. These type of effects are typically observed during the processing of the projections. In the present case, we did not observe a significant deformation of the sample along the several hours of acquisition, at least at length scales of ∼100 nm. This observation agrees with similar experiments performed on resin-embedded, osmium-stained Arabidopsis specimens (Polo et al., 2020).

FIB/SEM was performed as described previously (Knott et al., 2011). The retina sample from the wild-type mouse analyzed using PXCT or the retina in a trimmed Epon™ block (Fig. S2) were mounted on a regular SEM stub using conductive carbon and coated with 10 nm carbon by electron beam evaporation to render the sample conductive. Ion milling and image acquisition were performed simultaneously in an Auriga 40 Crossbeam system (Zeiss, Oberkochen, Germany) using the FIBICS Nanopatterning engine (Fibics Inc., Ottawa, Canada). Prior to starting the fine milling and imaging, a protective platinum layer of ∼300 nm was applied on top of the surface of the area of interest overlapping the OPL and ONL of the retina using the single gas injection system in the FIB/SEM. A large trench was milled at a current of 16 nA and 30 kV around the area of interest. For slice and view imaging, a gallium-ion beam of 600 pA at 30 kV and a cutting depth per slice of 20 nm (sample measured in PXCT) or 10 nm [osmium-thiocarbohydrazide-osmium (OTO) (Seligman et al., 1966) sample] was used. The OTO sample consists of a double osmium stain bridged by a reaction with thiocarbohydrazide. SEM images were acquired at 1.9 kV (60 µm aperture) using an in-lens ESB electron detector with a grid voltage of 500 V, a dwell time of 1 μs and a line averaging of 20 lines (OTO sample) or two lines (sample measured in PXCT). The pixel size was set to 5 nm and was tilt corrected.

FIB/SEM tomograms were analyzed using Imaris Image Analysis Software (Bitplane). Three-dimensional manual segmentation was performed using Avizo 3D Software (FEI). For the PXCT tomograms, the images were binned (2×2×2) and filtered with two iterations of anisotropic diffusion (Malik and Perona, 1990) before the manual segmentation.

We thank Marijana Samardzija (Grimm group, University of Zurich) for discussions and advice on retina degeneration and mouse models; Andrea Gubler (Grimm group, University of Zurich) for excellent technical support; and Susan Cohrs, Martin Behe and Cristina Müller from the Paul Scherrer Institute (PSI) for providing some of the mice that were used in this study. A part of the sample preparation was done at the Scientific Center for Optical and Electron Microscopy (ScopeM) at the ETH Zurich. We are especially thankful for the initial advice of Roger Wepf (Centre for Microscopy and Microanalysis, The University of Queensland, Australia) and training with Philippe Gasser, Falk Lucas and Joakim Reuteler (ScopeM). We thank Miriam Lucas (ScopeM), Johannes Ihli (Laboratory for Macromolecules and Bioimaging, PSI) and Sousan Abolhassani-Dadras (Nuclear Energy and Safety, PSI) for help with analyzing the data using Avizo software. The electron imaging part was performed with support of the Center for Microscopy and Image Analysis, University of Zurich, with the excellent technical assistance of Moritz Kirschmann. Thanks also to Gregor Cicchetti (PSI) for discussions and advice.

The peer review history is available online at https://journals.biologists.com/jcs/article-lookup/doi/10.1242/jcs.258561