A targeted 3D EM and correlative microscopy method using SEM array tomography

Imaging methods aid the analysis and understanding of biological processes. Advanced fluorescent light microscopy (FLM) techniques can answer questions about the localization, dynamics and interactions of proteins in various organs and cell structures. Genetically encoded fluorescence tools enable the study of live processes. However, many cellular processes cannot be fully visualized, because of the insufficient resolution of fluorescence labeling and the lack of visible reference structures within the cell. Much higher resolution of cellular components can be obtained with electron microscopy (EM), which is crucial for the analysis of cellular ultrastructure. Recently, cellular ultrastructure has been directly correlated with a fluorescence signal using correlative light and electron microscopy (CLEM; Caplan et al., 2011; Muller-Reichert and Verkade, 2014).

Serial section transmission electron microscopy (TEM), and tilt series combined with electron tomography, have been the primary tools for reconstructing ultrastructure within a defined volume. These methods have contributed significantly to our current understanding of the local 3D organization of biological samples (White et al., 1986; Bumbarger et al., 2007; Noske et al., 2008; Anderson et al., 2011; Saalfeld et al., 2012; Bock et al., 2011; Doroquez et al., 2014; Lee et al., 2016). Unfortunately, these procedures require a high level of training and are both time-consuming and prone to occasional loss of sections, resulting in loss of information. Recently developed alternative methods for serial sectioning, such as serial block face scanning electron microscopy (SBFSEM) and focused ion beam scanning electron microscopy (FIB-SEM), which use back-scattered electrons to image the block surface, have helped to address biological questions that require the analysis of large EM volumes (Denk and Horstmann, 2004; Heymann et al., 2006; Knott et al., 2011; Kizilyaprak et al., 2014; Titze and Genoud, 2016). During these approaches, the block surface is imaged, and then sections are removed. The new block surface is then imaged again. This cyclical repetition results in the generation of an automatically aligned stack of numerous sections. Although these approaches result in efficient 3D data acquisition, re-imaging the same sample at another region of interest or at a different resolution is impossible, as is on-section immunolabeling or post-staining contrast enhancement with heavy metals. These problems can be overcome, as has been shown in several studies that correlated light and FIB-SEM data on cells in culture, in brain tissue and in developing blood vessels in zebrafish (Bushby et al., 2012; Lucas et al., 2012, 2014; Maco et al., 2013; Bosch et al., 2015; Blazquez-Llorca et al., 2015). Notably, such elegant studies require rare and expensive equipment, available only in specialized labs with limited external access.

Array tomography (AT) was developed by Micheva and collaborators to study protein localization in synapses (Micheva and Smith, 2007; Micheva et al., 2010; Collman et al., 2015). The samples are embedded in a hydrophilic resin and sectioned into long ribbons of serial sections (the arrays) using an ultramicrotome. The sections are then transferred to a glass coverslip (instead of a microscope grid) and analyzed by light microscopy followed by scanning electron microscopy (SEM). To localize a large number of proteins in a single series of sections, several rounds of fluorescence labeling are performed, using antibodies combined with an antibody elution step before each additional round. The EM-like sample preparation procedure enables analysis of the cellular ultrastructure with the localization of a fluorescence signal.

AT is used in different procedures, depending on the research goal. (1) Multiplex FLM is used to detect numerous fluorophores (Table 1, Multiplex labeling). The mild sample treatment and hydrophilic resin embedding usually compromises the quality of the ultrastructure. (2) Serial EM is used for volume acquisition and region of interest (ROI) recognition, when ultrastructural preservation is crucial (Table 1, Volume reconstruction). As the back-scattered SEM mode is used, high contrast with multiple heavy metal stains is required. This preparation uses standard or enhanced contrast TEM with epoxy resin-based embedding. (3) Correlative fluorescence EM is used to preserve fluorescence and ultrastructure quality, both en bloc and on section (Table 1, CLEM). Freeze-substitution with acrylic resin embedding can provide efficient fluorescence retention (Nixon et al., 2009; Kukulski et al., 2011; Watanabe et al., 2011; Peddie et al., 2014; McDonald, 2014; Burette et al., 2015; Kremer et al., 2015; Paez-Segala et al., 2015; this study). However, the retention of the fluorescence signal with this approach is unpredictable and largely dependent on the specific fluorescent fusion protein used.

Although AT is a powerful technique, its use has been predominantly restricted to addressing questions in neurobiology, for which it was developed (Oberti et al., 2011). Recent studies have applied AT to other fields (Jahn et al., 2016; Markert et al., 2016). For AT, sections are cut and transferred to a solid support, such as a coverslip, an indium-tin oxide-coated coverslip, a silicon wafer or a Kapton plastic tape (Micheva and Smith, 2007; Pluk et al., 2009; Kasthuri et al., 2015) using special slide/wafer retraction mechanisms (Wacker and Schroeder, 2013; Hayworth et al., 2014). AT-SEM is increasingly used as an alternative to TEM serial section SBF and FIB-SEM imaging methods to analyze large surface areas in a non-destructive manner. It allows the direct 3D correlation of CLEM samples with both fluorescence and electron microscopy data. Unlike other volume sections transferred to a solid support, these samples can be stored, re-imaged multiple times, and even re-contrasted with heavy metals to improve the signal-to-noise ratio (Schwarz and Humbel, 2014).

Large volume data analysis is required to answer some biological questions. Other questions can be answered by analyzing limited volumes from numerous independent samples or events. To provide a method to answer these questions that is accessible to the wider scientific community, without technical limitations, we have modified and adapted the AT technique. Our method encompasses the entire procedure from sample preparation to data analysis. We transfer serial sections to glass coverslips or silicon wafers (hereafter referred to as ‘wafer’), label with heavy metals and use high-resolution volume FLM and EM. We have optimized sample preparation, configured a diamond knife to facilitate the transfer of sections to support and devised an efficient sectional screening strategy. This method can be used with a variety of model organisms. It does not require additional expensive equipment beyond that of most modern research facilities. To demonstrate our technique, we analyzed both Caenorhabditis elegans and Drosophila melanogaster.

When analyzing polarized samples and small model organisms, the relevant cellular structure, or ROI, can be located using the known morphology and anatomy of the tissue as landmarks. The principal problem with these complex samples is that only one ROI may be present in each sample. Because imaging only one individual structure would be insufficient to make a statistically significant observation, numerous samples must be analyzed.

Our strategy to efficiently localize the ROI in a single sample includes two-step flat embedding, precise block trimming (Fig. 1A,B; Kolotuev, 2014) and serial sectioning. We use light microscopy to locate the ROI rapidly for subsequent EM analysis, precisely orient the sample, and delimit the area for ultra-thin sectioning. To optimize the stability of the ribbon of consecutive sections, we remove the excess embedding resin surrounding the sample. With homogeneous tissue, excess embedding resin can be easily trimmed away from a sample (referred to as ‘tight trimming’). With non-homogeneous tissue, such as C. elegans or Drosophila, even ‘tight trimming’ cannot entirely remove the excess resin. Thus, we use our trimming diamond tool for ‘tight trimming’ to leave a minimal amount of excess resin around the sample. To acquire long ribbons of consecutive sections (>300 sections), we prefer to shape the surface of the block to a rectangle, with the edge parallel to the knife, such that the block length measures twice the length of the side edges. A larger sectional surface tends to produce more folds, whereas sections with a smaller surface tend to curl.

To produce horizontally aligned sections on support (glass coverslips or wafers), we use a modified diamond knife that greatly facilitates the efficient generation of long ribbons of consecutive sections (Fig. 1C-H; see detailed description below). Left intact, these long ribbons can complicate subsequent manual data acquisition: without obvious structural variation, mistakes will often be made during the monotonous analysis of long ribbons of consecutive sections. Thus, for our rapid screening approach, we carefully subdivide the long ribbons initially produced into shorter ribbons of consecutive sections (50-100 sections), and then horizontally align the shorter ribbons alongside each other, in rows, on the support. This way, ten short ribbons (i.e. more than 500 sections) can be easily aligned on a small support, be it a coverslip or a wafer (Fig. 1H). Then, instead of thoroughly analyzing every section within a long ribbon in a serial ‘vertical’ fashion (from 0 to 100+), we ‘leap’ between centrally located sections across the short horizontally aligned ribbons to screen visually for the ROI (see detailed example below).

Sequential section analysis using TEM grids is very tedious and numerous problems can emerge. If the sectional surface is relatively large, then very few sections will fit on one grid. If multiple sections do fit on one grid, then sample collection is error-prone because of the risk of sectional loss. Data analysis is also very time consuming because each section on the grid must be manually screened to find the ROI and orient the sample. For serial section collection, even a relatively short sequence of sections (e.g. 100) transferred to support (glass coverslips or wafers) presents significant advantages compared with sections on TEM grids (Fig. 1I,J). On support, 115 serial sections can be analyzed in less than 15 min after orienting the trimmed block optimally relative to the AT knife. Analysis of 115 sections on TEM grids would require much more time and result in significantly less precise data. For any scientific question that needs a resolution compatible with FIB-SEM and SBFSEM, our transfer technique has clear benefits over serial section TEM.

Both the histo and the ultra jumbo diamond knives (Diatome) hold a coverslip in their large boat at an optimal angle for collecting semi-thin sections. These knives can be used to generate arrays of sections, i.e. long ribbons of consecutive sections. In the published AT procedure, the array of sections is moved to the rear edge of the coverslip where it touches the interface between the glass and the water (Micheva and Smith, 2007). To attach the sections to support, water is drained from the side of the knife boat using a syringe (Micheva and Smith, 2007), or the coverslip is pulled out of the trough using a micromanipulator (Horstmann et al., 2012; Wacker et al., 2016). In our hands, this method compromises the control of water during draining. The resulting rapid drying of the sections on support can create wrinkles or folds in the sections and cause ROIs to be lost from view.

We have developed a water-draining system for the histo jumbo diamond knife, which reduces the incidence of sectional folding and wrinkling, and we have modified the diamond knife to better suit AT: (1) We replaced the diamond blade for the histo jumbo diamond knife for semi-thin sectioning (∼500 nm) with the diamond blade for the ultra 35° knife for cutting ultra-thin sections (∼50 nm). Ultra-thin sections have better quality than semi-thin sections: unlike semi-thin sections, ultra-thin AT sections with a thickness of 50-100 nm give better z-resolution and generate fewer folds. Ultra-thin sections also improve the quality of subsequent analysis, as they adhere better to the support and provide higher ultrastructural resolution. (2) We increased the size of the knife boat to completely accommodate a coverslip on a flat support on the bottom of the trough at a 0° angle. The support for the coverslip is surrounded with a ditch to allow for user-controlled water draining. A hole in the bottom of the boat facilitates the insertion of a flexible tube attached to a syringe (Fig. 1C,D). Using the syringe, water can be added to or retracted from the boat without creating turbulence, which had presented problems while using previous methods (Fig. 1E). To attach the sections to support, water is drained slowly through the hole in the bottom of the boat. Slow water removal avoids turbulence and allows the sections to dry slowly, without wrinkles and folds.

This technique of attaching the section ribbons to support is the same for all types of support. After generating the ribbon of the desired length (Fig. 1F), the water level in the boat is gradually lowered, by draining the water slowly through the hole in the bottom of the boat, either actively, by aspiration with the syringe, or passively, by allowing the water to drip through the tube without a syringe attached to it (Fig. 1G). When the water level reaches the height of the transfer support, the sections are manually aligned and arranged side by side (Fig. 1G,H, Movie 1). In our experience, this method greatly facilitates the ease and accuracy of section transfer for CLEM and volume EM. The same principle can also be used to collect histological sections for light microscopy analysis.

The immunolocalization of antigens is compromised by low z-resolution for both standard and super-resolution light microscopy. Also, the colocalization of multiple antigens using antibodies generated in the same species is often problematic. The advantage of AT is that sections can be cut as thin as 40 nm, and images of several cycles of immunofluorescence labeling can be acquired individually, and then superimposed. Physical ultra-thin sectioning can achieve a z-resolution of a few tens of nanometers, which is an order of magnitude better than the z-resolution of currently available super-resolution microscopes.

C. elegans is a transparent worm, which makes it an ideal model organism for light microscopy. Unfortunately, most specific primary antibodies available for C. elegans research are either mouse monoclonal or rabbit polyclonal. Thus, the colocalization of multiple antibodies on the same sections is often impossible. To demonstrate the power of the AT approach, we labeled serial sections of C. elegans larvae prepared by high-pressure freezing-freeze substitution (HPF-FS; see the Materials and Methods sample preparation section for details) on coverslip support. We stained/labeled the arrays of 100 nm thick sections with DAPI (nuclei, blue), anti-tubulin (red) and anti-ERM1 (EB1, green) in the first cycle and, after destaining, we stained for DAPI, anti-actin (red) and anti-VHA5 (green) in the second cycle (Fig. 2A,B). Both rounds of immunolabeling gave a strong signal, equally visible on individual sections and on the merge of all 25 sections (Fig. 2A,B, Movie 2). Before relabeling and re-analyzing the samples, we extensively washed the sections between the labeling cycles, and verified that the previous labeling had been completely removed. We used the manual acquisition of the images, improved their alignment by developing a new Fiji macro, and superimposed all markers in one data set (Fig. 2C). DAPI nuclear staining served as a useful reference marker to facilitate the alignment of the complementary signals provided by the two independent labeling cycles.

To monitor the quality and feasibility of EM acquisition, and to correlate the labeling data on sections after two labeling cycles, we imaged the corresponding ROIs with SEM using the back-scattered electron (BSE) detector. The fluorescence signal could be identified quickly, and the overall ultrastructure was well preserved (Fig. 2D). To improve the accuracy of superposition, we used DAPI nuclear staining and examined the nuclear ultrastructure and other structural features of the sample (Fig. 2E).

In complex tissues, drawing conclusions about the ultrastructure from a single section or a small number of sections can be impossible. Volume reconstruction of cells and tissues using the SBF and FIB-SEM methods is an efficient way to investigate the relationship between cells and organelles in their complex 3D cellular context. For both approaches, however, the ability to precisely target the ROI remains challenging and requires endogenous landmarks (Karreman et al., 2016). To illustrate the advantages of an AT-based solution in targeting the ROI, we imaged the Drosophila ovarian chamber at stage 3. Our previous efforts with the ovarian chamber had already exposed us to the difficulties of localizing the ROI in this tissue, and had actually prompted us to develop our current AT method that enables more direct sampling and screening.

Drosophila oogenesis is often used to model cellular processes, such as cell division, cell migration and stem cell differentiation (Horne-Badovinac and Bilder, 2005; Fuller and Spradling, 2007; McLaughlin and Bratu, 2015). During development of the ovarian chambers, the exterior follicular cells, the internal nurse cells and the oocyte undergo numerous structural changes. Nurse cells, which supply the necessary nutrients and cell building material to the oocyte, do not accomplish abscission but remain interconnected through the network of ring canals (RCs). During development and cell growth, the initially small RCs expand in diameter by about tenfold (Robinson et al., 1994; Hudson and Cooley, 2010; McLean and Cooley, 2013; Hudson et al., 2015). Imaging the RCs in their entirety using EM is difficult, mainly because of their large size and random distribution inside the ovarian chamber. On single sections, the RCs frequently appear as two spaced bars or brackets, and not as a circular structure, as in FLM micrographs. Based on our experience, localizing the RCs with FIB-SEM or SBFSEM tomography is problematic (Loyer et al., 2015).

We prepared the ovarian chambers using chemical fixation and flat embedding, sectioned them, and then transferred the ribbons of sections to wafers. Our fast screening strategy rapidly localized the RCs within the arrays of sections at low magnification. Then, we assessed the general orientation of the organism in a flat-embedded sample based on the anatomy of the sample (Fig. 3A). After sectioning, we localized the RCs based on known landmarks in the organism. For this, screening the sections on wafer at low magnification resulted in the rapid identification of regions to analyze at higher resolution (Fig. 3B,C). Initially, we collected ribbons of consecutive sections, aligned them side by side (Fig. 3B; each ribbon of sections is colored distinctly), and then tracked cellular features (or landmarks) along these ribbons. For faster screening, the sections were not screened in consecutive order but, rather, ‘horizontally’, by scanning the middle section from each ribbon for RCs (Fig. 3C; the color of the frame corresponds to the color of a given portion of the array, selected in proximity to the white arrow at the panel B). Once an RC was recognized and located (Fig. 3C, orange), we could easily track the entire relevant sequence of sections by sequentially following the sections along the ribbon, using the reference section as an anchor. This simplifies the screening logic: one starts from the beginning of the relevant sample portion and then collects sections for as long as necessary. In this way, the localization of anatomical features is more efficient, rapid and precise, avoiding the acquisition of unnecessary data sets.

To localize the four RCs, which connect the future ovary and the nurse cells, we applied the screening strategy described above. At low magnification, we identified RCs in sections and, as expected, the canals appeared only partially and did not resemble rings (Fig. 3D). After acquisition, the overlay of the sections allowed the modeling of the rings in their entirety, and the identification of the position of the ring on the cell surface (Movie 3). The same set of sections can be reused to reconstruct the interaction between additional pairs of cells in the same ovariole, or the sections can be re-imaged to analyze the structure of the RC in ovarioles of another developmental stage. Because the sections remain intact on the surface of the wafer, a ring of particular interest can be revisited using different acquisition parameters for more detailed study (Fig. 3E, Movie 4). The aligned data set can be further analyzed using different modeling and rendering programs, providing the precise identity of different cells, their spatial distribution and their interactions (Fig. 3F,G, Movie 5). The resulting data can be further mined by targeted analysis of relevant 3D data portions (Fig. 3H, Movie 5). Once the various stages of ovarian chambers and RC maturation have been collected and maintained on wafers, a library can be established. As in the case of ovarian chambers, any sample collected on wafers can be re-analyzed and re-imaged when new questions arise. Arrays can be stored indefinitely and, with proper annotation, can be used by several different research groups. In addition to the examples presented here, this strategy can be applied to answer similar questions in other complex tissues (Fig. S1).

We next extended our multiplex immunolabeling protocol and 3D screening with a reconstruction strategy, focusing on an organism bearing genetically encoded fluorescent markers. To retain the endogenous fluorescence en bloc for AT, we tested several preparation protocols. HPF-FS in acetone-containing uranyl acetate with subsequent embedding in HM20 methacrylate resin gave the most satisfactory results. With this, we could visualize various fluorescent constructs, both en bloc and on section (Fig. 4A-D), facilitating the localization of the ROIs in 3D. Moreover, searching for the fluorescence signal in a light microscope helped to localize the ROI faster than using EM screening for desired morphological features. We processed C. elegans larvae using this protocol. In addition to the endogenous signal, the sections could be labeled with antibodies, providing additional data (Fig. 4C). Here again, the superposition of the fluorescence and the EM images was reliable and straightforward, making the direct CLEM reconstruction and analysis precise and unambiguous (Fig. 4D).

To further facilitate rapid correlation of ROIs and automate the acquisition process, we used the MAPS software (FEI). This provided automatic acquisition of both fluorescence and SEM images. We then superimposed these images, allowing the transfer of the same ‘geographical’ parameters from one type of microscope to another (Loussert Fonta and Humbel, 2015). Using the CorrSight light microscope (FEI) and MAPS software, large regions can be acquired automatically at low magnification (×10), with the separate single images stitched together to form one large file. This provides an overview of the sections on the support, which is useful to identify the current location within the ribbons of consecutive sections. Imaging is then carried out using various objectives, imaging modes and several fluorescence channels to localize the proteins of interest. For correlation, high-magnification (×63, ×100) large-scale images of the fluorescent structures on the relevant sections can be acquired by tiling (Fig. 5A). After light microscopy, the project, containing the recorded light micrographs, can be transferred to a scanning electron microscope (Fig. 5B). Using anatomical landmarks, all light micrographs can be aligned and overlaid on the map of the electron micrograph and high-resolution electron micrographs can be recorded (Schwarz and Humbel, 2014).

To highlight the versatility of this technique, we analyzed the excretory canal of C. elegans. This single cell organ is located in the pharyngeal region of the worm and, during development, sprouts four tubular extensions that run the length of the entire body of the animal (Fig. 5C; Kolotuev et al., 2010, 2013; Sundaram and Buechner, 2016). Fast HPF-FS preparation enabled the preservation of the fluorescence vha-1::GFP signal (Fig. 5D). This greatly facilitated the recognition of the ROI after transfer to SEM, as the image has a similar appearance in both modalities.

The MAPs program can facilitate the automatic acquisition of both fluorescence and SEM images, subsequently superimpose these images, and provide a result analogous to that of the ‘Google Earth’ principle. First, images are acquired using low magnification (Fig. 5B) and then higher magnification is used (Fig. 5E,F). Then, ‘stitched images’ are composed of several single images at a given resolution so that both the overall field of view and the individual images of particular ROIs can be reused for future analysis. Compared with analyzing sections on TEM grids, using arrays permits the analysis of larger sections. Thus, more individual samples can be simultaneously collected (Fig. 5E, yellow; n>30). The fluorescence signal is easily detected on sections, facilitating identification of animals in the desired orientation and/or developmental stage. After the sample is identified with FLM, SEM data can be acquired either by manually correlating the position of the sections, or by using a MAPS-mediated site localization. After the ROI has been found on a reference section, one can select the desired resolution for the EM acquisition (Fig. 5G,H). Then the necessary ultrastructural data can be obtained by volume reconstruction. Within the tissue, the desired cell type can be ultrastructurally indistinguishable and/or difficult to localize. One example is the cell body of the early larval stage excretory canal, which is located in a densely packed environment and is not evident on routine ultra-thin sections. The fluorescence signal can aid in identifying a particular cell at the ultrastructural level (Fig. 5G). Once the fluorescence signal allows recognition of the cell of interest, EM analysis of the subsequent sections no longer requires correlation with the fluorescence data: the position of the cell on subsequent sections becomes evident (Fig. 5H).

The adult excretory canal extension is 100-1000 µm long and 50-100 nm in diameter, making it difficult to image the overall volume of the canal arms using SEM. Only very rarely does the fixed animal remain straight after processing, so orienting for longitudinal sectioning is a tedious procedure. Usually, only a small portion of the body can be correctly oriented. Locating an ROI may require screening and acquisition of hundreds of carefully collected serial sections (Fig. 6A). Thus, it is complicated to orient precisely these extensions for longitudinal sectioning. The specimen often curves after fixation, and the recognition of the defragmented structure without the overall body context is often difficult, as the same structure appears different, and in a different location, depending on the sectioning axis (Fig. 6B). In this case, the fluorescence serves as a landmark to permit the easy recognition of the cells of interest. For this purpose, we labeled the excretory cell in green and a random cell in its vicinity in blue. Superficially, these cells appeared identical, and if one had to distinguish them by a ‘blind search’, it might be impossible.

Here again, the combination of fluorescence-mediated screening with subsequent targeted EM acquisition facilitates precise orientation. First, the appropriate fluorescence image is found on section. It is useful to label the sections with a nuclear marker (e.g. DAPI, Draq5) to assess the orientation of the sample and to avoid confusion in repositioning the fluorescence signal on the EM data. Second, fluorescence data rapidly guides the analysis to the relevant ROI, after which the required number of sections or the desired volume of data is collected (Fig. 6C). In our example, the sections are spaced less than 200 nm apart, and the portion of the canal that connects the extension to the cell body is no longer seen from one section to another (Fig. 6D, arrow). Without fluorescence, the detached part would not have been evident using EM data only (Fig. 6, arrows).

Our results demonstrate that when the endogenous fluorescence signal is preserved throughout the preparation, the random-sectioning approach facilitates identification of randomly oriented samples. When combined with the MAPS correlation software, the structures of interest can be rapidly identified and imaged at high resolution by EM, enabling fast-throughput analysis of a statistically significant number of organisms.

This study presents an efficient AT procedure to facilitate the analysis of a single ROI in a whole organism on serial sections that is applicable for numerous biological samples. On serial sections, the ROI can be labeled with a fluorescent tag and correlated with the corresponding electron micrograph. The serial sections can then be quickly screened using fluorescence labeling or directly using SEM.

Besides the SEM, our method does not require additional expensive equipment for an EM lab or facility. The newly designed diamond knife we introduce can be used with any ultramicrotome, and its application is technically easier than performing classical ultra-thin sectioning. In contrast, the approach using SBFSEM with FIB-SEM has a very high cost and limited accessibility. An alternative approach using the ATUMtome machine for automatic section collection from neuronal samples was recently demonstrated (Kasthuri et al., 2015). Likewise, an alternative AutoCUTS method was recently used to collect sections of C. elegans larvae (Li et al., 2017). However, this time-consuming and more expensive strategy is only worthwhile when there is a requirement for thousands of sections, as is often needed for brain samples. In contrast, our direct AT method requires a modified diamond knife, whereas the ATUMtome strategy requires an expensive and specialized machine (available at selected facilities only).

For both the ATUMtome and our direct AT method, libraries of serial sections can be stored indefinitely and re-analyzed to address future research questions. For studies involving multiple organisms or organs, for which neither serial section TEM nor volume EM approaches would be efficient, our screening method provides an accessible solution. For questions that can best be answered by analyzing many different samples (e.g. wild type versus mutants), libraries of AT sections can be extremely useful. A single 2×4 cm wafer can contain all the necessary information to permit the analysis of a small organism quickly. Also, the researcher need not immediately acquire the entire sectional area at high resolution, as is required for block-face microscopy. High-resolution collection of only the relevant data results in greater precision and shorter acquisition time. Fewer electronic data are stored, as information that is irrelevant to the initial question is not collected. Thus, statistically significant EM volume data can be acquired in a timely fashion. Further, compared with block-face methods, the sections generated by AT can first be imaged at low magnification for orientation, and then the ROI can be imaged at higher resolution for a more detailed analysis. Direct CLEM, with or without automation, can provide an additional level of volume correlation and can simplify the EM identification of similar structures. The fields of cell and developmental biology, model organism analysis, and even clinical research could benefit from this technique, to address such mechanisms as cell migration, tissue remodeling, and the fine structure of cell division and abscission (Fig. S1).

Most of the technical problems with the use of the existing SEM imaging fields and modalities relate to the lack of precision of automatic acquisition. Current software systems that support AT include Atlas, with Zeiss scanning electron microscopes, and MAPS, with FEI electron microscopes. With both software systems still in development, future improvements could facilitate the automation of robust and efficient serial section acquisition; for example, the acquisition process could be automated, and more solid and user-friendly data alignment algorithms could be provided.

In conclusion, our AT method can be used to resolve long-standing questions in cell and tissue biology that require 3D ultrastructural analysis, and to localize antigens in defined structures. This approach could be useful for both biological research and clinical analysis, as high-quality results can be efficiently achieved using equipment that is, to a large extent, already present in many imaging facilities.

The samples were either chemically fixed or processed using the HPF-FS method (Table 1).

The samples were fixed for 2 h at ambient temperature in 2% (w/v) formaldehyde and 2.5% (w/v) glutaraldehyde dissolved in 0.1 M phosphate buffer, pH 7.4 (EMS, 19340-72, 11650, 19344-10, respectively). They were then post-fixed in 1% (w/v) osmium tetroxide (EMS, 19152) and 1.5% (w/v) potassium ferricyanide in cacodylate buffer (EMS, 26604 1A) for 1 h at ambient temperature, followed by 1 h incubation in 1% (w/v) tannic acid (Sigma-Aldrich, 403040). After extensive washes in water, the samples were incubated in 2% osmium tetroxide followed by 1 h in 1% (w/v) uranyl acetate in water and dehydrated in an ascending series of ethanol followed by an ascending concentration of Epon (EMS, 1420) in ethanol mixtures (Starborg et al., 2013; Kolotuev, 2014). Finally, the samples were flat embedded using a two-step procedure, and then polymerized at 60°C for 48 h as previously described (Fig. 1A, Table 1; Kolotuev, 2014).

HPF-FS was performed using a modified version of the previously described method (McDonald, 2014). Carriers containing HPF-immobilized samples were transferred to vials containing freeze substitution media [0.1% (w/v) uranyl acetate, 2% (v/v) H₂O in acetone] and then placed in the freeze-substitution unit (AFS2, Leica Microsystems), cooled to −140°C, and heated to −50°C over 8 h. The samples were then infiltrated by ethanol and resin mixtures. After freeze-substitution, the samples were flat embedded, either in Epon embed 812 resin (EMS, 1420 morphology analysis), or in the monostep HM20 hydrophobic methacrylate resin (EMS, 14340) for the samples containing fluorescently tagged proteins, or those to be used for immunofluorescent labeling (Table 1; Kolotuev, 2014).

Blocks were trimmed to ∼100×200 µm cuboids using a trim 90-diamond blade (Diatome). Sample block sides parallel to the knife were coated with a mixture of one part Weldwood contact cement (Néoprène Liquide, Sader) to three parts xylene. This mixture increased the adhesiveness of the sections to each other to produce a long ribbon of consecutive sections (>300). The resulting long ribbons of consecutive sections were subdivided into shorter ribbons (50-100 sections) and then aligned on support, side by side (Kolotuev, 2014).

To generate planar arrays of sections on support, we used a modified ultra 35° diamond knife with a custom-built boat (Fig. 1C,D; Diatome; see Results for details) mounted on a UC7 ultramicrotome (Fig. 1E; Leica Microsystems). Ribbons of consecutive sections were generated with a length that varied according to the eventual goal of the AT procedure: 50-200 sections per ribbon, 50-150 nm thick. Each array of ribbons aimed for multi-labeling was transferred to a glass coverslip (VWR, 631-1575), or, if it was dedicated for EM reconstructions and CLEM, to a silicon wafer (Ted Pella, 16015). Before use, glass coverslips were cleaned and coated with 0.5% gelatin solution (EMS, 16564) added to 0.05% chromium potassium sulfate (Sigma-Aldrich, 243361), following a previously described procedure (Micheva et al., 2010). We followed the manufacturer's instructions to obtain a 1-2 nm carbon layer for the samples collected on the glass slide using CCU-0101 LV machine (Safematic). A thicker layer can compromise the quality of the SEM images. Image collection on the ITO-coated slides should eliminate the need in this step (I.K. and A.M., unpublished observation,).

Round wafers were cut into 2×4 cm rectangles using glass breaking pliers 6″ (Ted Pella, 7295), cleaned, and glow discharged for 30 s using the current plasma HT of 15 mA (Cressington 208, Cressington Scientific Instruments). We did not carbon coat the samples on wafers because we did not detect any charging effect during the image acquisition.

Sectioning problems can compromise both the quality of the sample and the reliability of the subsequent analysis. The appearance of wrinkles on the surface of sections can ruin the entire AT data set. The impact of a fold on data analysis is difficult to quantify, mainly because only folds that affect the ROI are deleterious.

We distinguished two types of wrinkles that can appear on the surface of sections: (1) folds, causing global damage to the section, and (2) microfolds, small wrinkles in limited areas of the sample surface. We could detect both types of wrinkles on sections, but frequently they affected only the excess embedding resin surrounding the sample. Also, on several sections, the wrinkles could be found on the sample surface, without affecting the ROI (Fig. S2). Even in very extreme cases of severe wrinkles, it is difficult to interpret their impact on data analysis. In Fig. S3, four independent samples of a section are shown. Even if samples 1, 2, or 3 have large folds or debris, this will not affect the data analysis. However, if sample 4 had a fold in the bottom half, it would only be ruined if we were interested in its bottom half. If we wanted to analyze its upper half, then the wrinkle would not affect our analysis. The situation with microfolds is very similar if they happen to occur within the sample but not within the ROI (Fig. S3). Thus, we decided not to quantify the number of folds but, rather, to characterize the conditions under which they appear and possible ways to prevent them.

Below we list factors that can lead to folds and wrinkles and how these can be minimized.

The size of the sample surface appears to be essential: smaller sections result in fewer wrinkles because there is less surface to generate wrinkles. Further, density differences within the sample can influence the quality of sectioning. Epoxy resin will create a copolymer within the biological sample, which has a different density than the empty resin. Thus, excess embedding resin should be removed. Within homogeneous tissue, all excess resin is removed, leaving only a rectangular block of embedded tissue. For smaller non-homogeneous samples, such as C. elegans or Drosophila, it is difficult or impossible to remove the entire resin around the sample. Thus, we advise maintaining a similar amount of resin on each side of the sample.

For better stability of the ribbon of consecutive sections, it is important to trim the block as close to the sample as possible, shaping its surface into a rectangle, with the edge touching the knife measuring twice the width of the edges at the side. A larger section surface produces more folds, whereas sections with a smaller surface tend to curve more. Curving of the ribbon of consecutive sections (section curling) is problematic, especially for ribbons that will be aligned side by side on support. With some experience, this problem can be overcome (Movie 1).

This parameter is as important as the block edge size and has a significant effect on the quality of the resulting sections. Ultra-thin (50-80 nm) sections attach better to the support and have fewer microfolds compared with semi-thin sections (100-250 nm).

We found the rate of water retraction to be crucial for successful transfer of the ribbons of consecutive sections to the slide support. The support surface should be hydrophilic. Excessive hydrophobicity of the support can cause folds in the sections. Thus, supports dedicated for SEM analysis (wafers, glass slides or Kapton tape) should be carefully pretreated using a standard glow discharge procedure. Even with a very disorganized collection of sections, if the process of drying is optimal, there is a lower chance of the appearance of folds (Fig. S4, can be zoomed in) and the development of microfolds in non-homogeneous samples is much less likely to occur.

Occasional debris on the section surface can be difficult to eliminate. To reduce the occurrence of debris, the knife and the basin must be meticulously rinsed with distilled water filtered with a 0.22 µm syringe filter. Coated glass slides must be carefully prepared, and wafers thoroughly cleaned. Water filtering and knife cleaning before the generation of the sections can help to minimize the accumulation of dust and debris on the surface of the sections.

For multiplex fluorescent labeling, the arrays were labeled with a subset of different antibodies using a previously described procedure (Micheva et al., 2010). Briefly, ribbons of consecutive sections on a glass coverslip were incubated in a humidity chamber with 50 mM glycine solution in TBS (Sigma-Aldrich, T5030) for 5 min, then with blocking solution containing 0.1% BSAc (Aurion, 900-099) in TBS for 5 min. Specific primary antibodies mouse anti-actin (Cedarlane, CLT 9001), mouse anti-tubulin (Invitrogen, 322500), rabbit anti-VHA5 (courtesy of Dr Labouesse, Institut de Biologie, Paris, France) and rabbit anti-EB1 (Santa Cruz, SC 15347) were diluted to 1:100 v/v in TBS and incubated overnight at 4°C (Nicolle et al., 2015). Secondary antibodies Cy2 anti-mouse (Abcam, Ab-6944) and Cy3 anti-rabbit (Abcam, Ab-6939) were diluted to 1:100 v/v in TBS and incubated at ambient temperature for 2 h.

Samples were mounted using the Vectashield Antifade Mounting Medium (Vector Laboratories, H1200) and imaged the same day. Destaining and relabeling of the arrays was performed according to the experimental requirements (Micheva and Smith, 2007). For removing the bound antibody complex, a mixture of 0.2 M NaOH with 0.02% SDS was applied to the samples for 20 min, followed by extensive washing cycles. Before relabeling using the same method as the initial labeling, the samples were dried in an oven at 60°C for 20 min.

After completion of fluorescence image acquisition, the samples were extensively washed in double-distilled water, dried and contrasted for 15 min in a 4% uranyl acetate solution followed by a 10 min incubation in Reynolds’ lead citrate solution.

For SEM acquisition, samples prepared using the high-contrast procedure were directly observed without further manipulation. For all other acquisitions, the sample contrast was increased using the same parameters as for CLEM experiments.

Multiple fluorophores and fluorescence preservation procedures for CLEM have been developed (Nixon et al., 2009; Kukulski et al., 2011; Watanabe et al., 2011; Peddie et al., 2014; McDonald, 2014; Burette et al., 2015; Johnson et al., 2015; Kopek et al., 2017). Although efficient for many types of proteins, the level of fluorescence of genetically encoded markers on sections varies from construct to construct. Although we did our best, we could not predict under which preparation conditions it would be preserved at all, and to what extent. For some constructs, fluorescence remains visible inside the block and on sections even without using an anti-bleaching solution. For other samples, even if we were not able to detect a signal inside the block the fluorescence was often visible in the sections. For these samples, we could preserve the fluorescence signal for at least 1 year after the preparation. For some samples, however, although the fluorescence signal was initially present, we were not able to preserve the fluorescence in blocks or on section. A systematic analysis of fluorescence preservation as a function of a protein type (transmembrane, cytosolic, secretory, etc.) could further facilitate our understanding of the problem of fluorescence preservation. In the future, we hope to combine our observations with those of other groups that have experience with fluorescence retention using other species and fluorescent constructs to understand the factors influencing fluorescence retention and to develop a robust preparation protocol.

Intuitively, thicker sections should provide a better fluorescence signal when genetically encoded markers are used. However, there is a thickness-to-stickiness relation, in addition to the loss of EM details in thicker sections: thinner sections had better adherence to the surface. Also, photo-bleaching was even more of a problem. Thus, to find a reliable way to quantify the ‘signal emission’ from sections of different thickness might require complicated math/physics calculations, which would likely not be applicable for all types of fluorescent constructs. However, the overall preservation of sectional fluorescence is a concern because, after fixation, the signal is not linear and varies with the type of construct.

Wide-field fluorescence image acquisition was made using a Leica DMIRB inverted microscope or a Leica DMRXA upright microscope equipped with a 20×/0.4 NA plan, a 63×/0.7 NA PL Fluotar or 100×/0.7 air and 1.4 NA PL APO oil objectives, and a Coolsnap HQ2 CCD camera (Photometrics).

Confocal fluorescence images were taken using a TCS SP8-SMD confocal microscope equipped with both 63×/1.4 NA Oil HC PL APO and 40×/1.3 NA Oil HC PL APO objectives (Leica Microsystems). Fluorescence signals were collected sequentially through photon-counting hybrid detectors (Leica) to increase the signal-to-noise ratio.

For light microscopy tile acquisitions, the CorrSight light microscope running the MAPS software was used (Loussert Fonta and Humbel, 2015). The microscope was equipped with Zeiss objectives: 5× NA 0.15, 20× NA 0.8, 40× NA 0.9 and 40× NA 1.3 oil immersion; and DAPI, 488 nm and mCherry filters. The light micrographs, bright-field and fluorescence images were recorded with MAPS and stored as an aligned package in a project.

For SEM acquisition, we used a JEM 7800F with retractable backscattered electron detector (JEOL) with the following parameters: accelerating voltage, 8 kV; probe current, 1.6 nA; WD, 5 mm; dwell time, 10 µs/pix; frame size, 4096×4096. For MAPS-related experiments, Quanta FEG 250 SEM (FEI) was used with the following parameters: accelerating voltage, 10 kV; probe current: 0.8 nA; WD, 7 mm; dwell time, 30 µs/pix; frame size, 4096×4096. The MAPS projects were transferred to the Helios 650 FIB-SEM (ThermoFisher). A reference image was taken with SEM. Light micrographs were then aligned to this image and used to find the ROI to be analyzed at high resolution using SEM (Loussert Fonta and Humbel, 2015).

To align the fluorescent stacks of sections, we developed the computational volumetric analysis protocol, based on Fiji and ICY programs (Schindelin et al., 2012; de Chaumont et al., 2012). The coregistration procedure will depend on the data set. Using Fiji software, we selected C. elegans embryos in each stack using an automatic threshold, and signals located outside of the chosen area were deleted. The first two steps can be done automatically using a rigid registration. Indeed, as images come from the same array with a microscope always using the same settings, the displacements between two images are rotation and translation. In the image processing community, numerous algorithms aim to solve the rigid registration task. We used the ‘Register Virtual Stack Slices’ feature available on Fiji (imagej.net/Register_Virtual_Stack_Slices). This method uses the SIFT feature to find corresponding regions in image pairs. To facilitate the use of this registration algorithm for array data, we developed a Fiji macro that runs in three steps. (1) ‘Split Stack’: this step splits the array hyperstack into various image sequences, one sequence for each staining of the FLM data. The sequences are stored in output subfolders. The operator need only specify the number of channels (i.e. staining) in the stack. (2) ‘Register Reference’: this step calculates the registration transformation matrices using the reference staining. Usually, the DAPI staining is used as a reference for FLM. The operator needs only to provide the reference channel directory, and the output directory where the transformation matrices will be stored. This directory is already created by the macro, the operator can simply select it. The algorithm settings are: Reference channel, channel used to estimate the transformation; Registered image size, size of the output image canvas. It should be larger than the original image for all of the images that have been transformed to fit into the canvas. (3) ‘Apply transformation’: this step applies the transformations calculated in (2), to all of the staining series.

The co-registration procedure will depend on the data set. Depending on the image content, the registration algorithm may fail to locate any corresponding regions between the separate array images. This can be resolved using pre-filtering of the reference staining before registration (i.e. between step 1 and step 2). The operator can then apply a filter to enhance the brightness, the contrast, or the gradient of the array. This optional step had not been implemented in the macro because it is application dependent. The use of the macro is similar to registering light microscopy and EM array, but for the EM alignment, a single channel is used. For the complete macro, please see supplementary material.

EM/FLM registration cannot be done automatically because the resolution and the content of images are different. Thus, a supervised registration method is required. For this, landmarks that match the reference regions on FLM and EM micrographs are used with the ICY plugin ec-CLEM (Paul-Gilloteaux et al., 2017). We coregistered arrays obtained with the macro described above and registered the first image of the FLM array and the first image of the EM array with ec-CLEM. A detailed tutorial of ec-CLEM is available at icy.bioimageanalysis.org/plugin/ec-CLEM. ec-CLEM generates a file containing the registration transformation matrix. The operator can then apply this transformation to all of the FLM arrays with ec-CLEM.

For the SEM data alignment and 3D modeling, we used IMOD (University of Colorado), Fiji (LOCI), and Photoshop (Adobe) software (Schindelin et al., 2012; Kremer et al., 1996). To achieve a TEM-like appearance, the SEM micrographs were contrast-inverted using IMOD or Photoshop.

We thank the following colleagues for their help: Kristina Micheva, Gareth Griffiths, Lucy O'Brien, Till Matzat, Paola Moreno-Roman and Karen Lunde for critical reading of the manuscript and fruitful discussions. We thank Loic Joanny (SCANMAT, Université de Rennes 1) for help with the SEM, Jean Daraspe for sharing the 3Dmod script, Willy Blanchard for the macro imaging and Tom Garber for the help in 3D image rendering.