Macromolecules are essential to life. Determining their structure is crucial to understand how they perform their myriad biological functions. Recent advances in cryogenic electron microscopy (cryo-EM)—also known as the resolution revolution1—have made it possible to routinely achieve structures of isolated proteins and other macromolecules at near-atomic, and even atomic, resolution2,3. However, isolating macromolecules from their biological context can be challenging, or even impossible. More importantly, macromolecular conformations and functions are deeply rooted in their cellular context, and isolating macromolecules from their native surroundings can disrupt intricate interaction networks and change macromolecular structure. It is therefore essential to study macromolecules in their unperturbed cellular environment, to understand their function and conformational landscape in relation to interaction partners, cellular location and the physiological state of the cell.

Recent advances in cellular cryogenic electron tomography (cryo-ET) allow scientists to perform these critical studies. To preserve the fine molecular details of the cell in a near-native state, cells, tissues or small organisms are first frozen within milliseconds to at least below −150 °C (ref. 4). Next, a cryo-ET instrument is used to acquire images at different tilt angles, which are subsequently used to reconstruct a volume of the sample to provide three-dimensional (3D) data on cells under near-native conditions. This can be done directly on thin samples (<500 nm) such as thin bacteria and the edges of mammalian cells5,6. Macromolecular structures can be determined from these data by subtomogram averaging, where hundreds to many thousands of 3D images are iteratively aligned and averaged7,8.

However, most biological samples are thicker than 500 nm, and two techniques have been developed to generate thin biological samples while maintaining their vitreous state: cryogenic sectioning and cryogenic focused ion beam (cryo-FIB) milling. Cryogenic sectioning uses a diamond knife to prepare <70 nm sections of cellular material9,10, but causes several artifacts, including severe compression in the cutting direction (Fig. 1a–c) and wrinkling of the sections11. As an alternative to cryogenic sectioning, cryo-FIB milling causes fewer artifacts (for example curtaining or redeposition), most of which can be avoided12,13,14,15 (Fig. 1d–l). The cryo-FIB milling process and its artifacts are explained in detail in refs. 14,15.

Fig. 1: Overview of vitreous cryogenic sectioning and FIB fabrication of lamellae.
figure 1

a, TEM low-magnification overview of a vitreous cryogenic section of high-pressure-frozen bacteria (Yersinia enterocolitica) on a holey carbon electron microscopy grid. Scale bar, 150 µm. b, Intermediate-magnification low-dose TEM overview of a vitreous cryogenic section. Knife marks created during the sectioning can be seen as parallel lines that run in the cutting direction of the diamond knife (black arrow). Several bacteria (Bac.) are visible in the holes and on the carbon. Scale bar, 500 nm. c, Tomographic slice recorded on a vitreous section of high-pressure-frozen Y. enterocolitica. Deformations of the bacterial membrane are visible (white arrows), which are common artifacts caused by vitreous sectioning. Scale bar, 100 nm. d, SEM overview of an electron microscopy grid with adherent dendritic cells infected with Y. enterocolitica, used to select cells suitable for FIB fabrication of lamellae. The enlarged area shows a cluster of cells suitable for FIB milling. Scale bar, 500 µm. e, FIB image of a lamella of ~1 µm thickness during the FIB fabrication process. All of the cellular material above and below a section of the cell is removed with the FIB at a shallow angle relative to the grid with increasingly lower ion beam currents, ultimately retaining a 50–300 nm thin lamella. The approximate directions of the ion and electron beam are indicated with arrows (green and red, respectively). The black arrow indicates the approximate viewing direction in f. Scale bar, 20 µm. f, Schematic side view of FIB milling geometry on a cell on a grid. The angles of the FIB and the SEM beam are indicated in green and red, respectively. The dashed line in the cell indicates where the lamella is formed after FIB milling. g, SEM image of the final ~200 nm thin lamella. Micro-expansion joints69 are visible on both sides of the lamella (as indicated by asterisks), which are formed with the FIB to prevent the lamella from cracking due to a build up of tension in the grid foil. Scale bar, 10 µm. h, FIB image of the lamella in g from the direction of the FIB during lamella fabrication. Scale bar, 10 µm. i, Low-dose stitched tilescan TEM overview of a lamella used to select positions for recording tilt series. The image shows the gold grid foil (red asterisk) and an organometallic platinum layer to protect the front edge from the ion beam (white asterisk). Scale bar, 5 µm. j, Enlarged area of the white dashed box in i. Several intracellular bacteria are visible. Uneven milling speeds, caused by differences in local densities of the cell, can create streak artifacts in the direction of the ion beam (black arrow) known as curtaining (white arrow). Ice contamination is indicated with a green arrow. Scale bar, 500 nm. k, Tomographic slice recorded on a FIB-fabricated lamella of Y. enterocolitica during intracellular infection of a dendritic cell. Several different macromolecule complexes are visible, including the needle of the bacterial type III secretion system (black arrow) and ribosomes (green arrow). 10 nm nanogold fiducials, used for tilt-series alignment, were computationally subtracted from the tomogram (red circle). Scale bar, 100 nm. l, Tomographic slice of a major histocompatibility complex class II compartment recorded on a FIB-fabricated lamella of a primary human dendritic cell. Scale bar, 200 nm.

FIB fabrication of lamellae originated from materials science, and pioneering experiments established the proof of principle that it can also be applied under cryogenic conditions to frozen biological samples without devitrifying them12,16,17,18,19. FIB fabrication removes all of the material above and below a section of the cell (Fig. 1e–h), leaving a thin (50–300 nm local thickness) slice of cellular material, termed a lamella. More recently, cryo-ET imaging on FIB-fabricated lamellae has been used to successfully study the molecular environment of the cell20,21 and is now rapidly evolving into a powerful in situ structural imaging technique.

This Review explores the unique strengths and weaknesses of cryo-ET on FIB-fabricated lamellae and which types of research questions it can answer by describing recent studies that use these techniques. Next, we outline the most recent advances in FIB preparation of lamellae, cryo-ET, data processing and subtomogram averaging that have made these studies possible. We then discuss the challenges in finding macromolecules of interest in the crowded cellular environment, and the role and potential of correlative light and electron microscopy (CLEM) for FIB-fabricated lamellae. Finally, we discuss the future potential of cryo-ET for in situ structural biology and which advances would facilitate its realization.

In situ structural studies of the cell by cryo-ET on FIB-fabricated lamellae

Cellular cryo-ET has several key advantages over other structural techniques, stemming from observing macromolecules within their native cellular environment. Besides structural findings, high-resolution imaging of the near-native cellular environment allows for visualizing cellular morphology in close detail.

Determining structures within the cell

Numerous atomic models of biomolecules based on high-resolution experimental data have been determined using structural techniques such as X-ray crystallography, single-particle cryo-EM and NMR. Cryo-ET on FIB-fabricated lamellae can complement these structural data in the near-native cellular environment, where all natively present (macro)molecular interactions are retained. This advantage makes cryo-ET a powerful method with which to validate (near-)atomic-resolution structures of isolated macromolecules obtained using other structural techniques. An example is the coat structure of coat protein complex I-coated vesicles, where cryo-ET on FIB-fabricated lamellae was used to validate a previously determined structure obtained on reconstituted vesicles22,23. Compared with in vitro reconstituted vesicles, the intracellular view uniquely shows their native architecture and the relative abundance of the coated and uncoated states, which provides insight about the relative half-life of coat protein complex I-coated vesicles22. Another example is the in situ structure of the proteinaceous shell formed around the replicative compartment of jumbo phages during infection of bacteria24. An intracellular subtomogram-averaged structure of the shell protein chimallin was limited to 24 Å, due to the variable curvatures and heterogenous nature of the shell. By fitting the 3.1 Å single-particle structure of the isolated chimalin monomer into the subtomogram averaging map, the higher-order native assembly of the chimalin shell could be determined, revealing a flexible geometry that can accommodate the variable curvatures of the replicate compartment.

Other examples of in situ structures include the transmembrane ribosome–translocon complex bound to the endoplasmic reticulum25,26; the retromer coat in algae27; the nuclear pore complex in algae, yeast and human cells28,29,30; the calcium channel CatSper in sperm flagella31; the centriole in algae32; the severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) spike protein during intracellular infection33; and the bacterial type III secretion system during intracellular infection of primary immune cells34.

Although most structures currently obtained with subtomogram averaging on lamellae have a resolution of between 1 and 4 nm, several recent studies have obtained structures at subnanometer resolution35,36. In two studies by Wang et al.37,38, structures were obtained of the thin and thick filaments in sarcomeres in muscle tissue, including that of nebulin bound to actin–myosin complexes at 5–10 Å resolution (Fig. 2a). Importantly, the structure of the molecular ruler nebulin was found to be very different within tissue compared with isolated structures, and different structural states were observed for the double heads of the molecular motor protein myosin.

Fig. 2: Examples of structures obtained within cells with cryo-ET on FIB-fabricated lamellae and subsequent subtomogram averaging.
figure 2

a, Left, tomographic slice of a sarcomere A band from skeletal mouse muscle tissue, obtained with cryo-ET on a cryo-FIB-fabricated lamella. Thin actin and thick myosin filaments are readily visible, as well as myosin double heads. Scale bar, 20 nm. Right, density map of the thin filaments obtained with subtomograms, averaging at a resolution of 5–7 Å. This density map enabled structural analysis of the structures of actin (green), tropomyosin (blue) and nebulin (magenta), as well as the structure of myosin double heads (heavy chain, yellow; essential light chain, orange; regulatory light chain, red). Adapted with permission from ref. 38, AAAS. b, Structure of the coronavirus pore complex in mammalian cells. Top left and middle, 7-nm-thick tomographic slices of double-membrane vesicles (DMV) formed by MHV and SARS-CoV-2, respectively. Pore complexes formed by coronaviruses span the DMV membranes (white arrowheads and inset on top right) and are probably involved in transporting viral RNA into the cytosol of the host cell. Scale bars, 50 nm. Top right, central slice of the subtomogram average of the pore complex from tomograms on FIB-fabricated lamellae of cells infected with MHV. Scale bar, 10 nm. Bottom, isosurface rendering of the subtomogram-averaged structure of the pore complex (orange) spanning the inner and outer DMV membranes (blue and yellow, respectively). The pore structure has a sixfold rotational symmetry and a central channel with a minimum diameter of ~2–3 nm. Adapted with permission from ref. 40, AAAS.

Visualizing the native cellular environment may also lead to the discovery of new binding partners39,40, or even unknown macromolecular complexes and structures. An example is the discovery of a coronavirus pore structure formed by the murine hepatitis virus (MHV) in FIB-fabricated lamellae of infected mammalian cells (Fig. 2b)40. This pore structure spans the double membrane of the MHV replication compartments, termed double-membrane vesicles, and may function as a transport gate for viral RNA to enter the cytosol. These molecular pores were also confirmed to be formed by SARS-CoV-2 in samples chemically fixed before ultrarapid freezing. Other examples of novel structures identified with cryo-ET on FIB-fabricated lamella include transmembrane nanobristles on intestinal microvilli in Caenorhabditis elegans41 and a dome-shaped protein complex with a still unknown function on the outer membranes of lamellar bodies in primary human small airway epithelial cells42. Knockout mutants39 or fusion proteins40 can be used to confirm the identity of proteins. To understand macromolecular structures in relation to function or disease, cellular stimuli and mutants can be used. For example, Weiss and colleagues39 applied cell stress stimuli before ultrarapid freezing of cyanobacteria to determine the structure of a novel protein complex termed septal junctions, in both open and closed conformations, using cryo-ET on FIB-fabricated lamellae. These junctions connect neighboring cells similarly to gap junctions in mammals.

Mapping structural heterogeneity within the cell

Macromolecular structures are seldom static and may change depending on subcellular localization or binding partners. Thus, determining the structures of macromolecular complexes within the cell allows the subcellular localization and spacing of macromolecules relative to one another and to organelles to be preserved. Cryo-ET on FIB-fabricated lamellae can be used to create a visual map of macromolecules in different conformational states by combining automated detection of macromolecules (for example, template matching) with subtomogram averaging and classification. Macromolecule identity, orientation, structural state and binding partners can be determined and mapped back into the tomographic volume. In this way, structural information is combined with macromolecular spatial organization and subcellular localization20,43. Combining structural states, macromolecular orientation and subcellular position to create a molecular atlas of the whole proteome has been termed visual proteomics43. Currently, subtomogram averaging is limited to macromolecules that are a part of larger macromolecular complexes (>300 kDa or thereabouts), because their coordinates in tomograms need to be (visually) determined (as discussed in the section ‘Interpreting tomograms of the crowded cellular environment’). This limitation greatly restricts the fraction of the proteome that can be structurally analyzed with subtomogram averaging and subsequently mapped back into the tomographic volume. Recent examples of studies that effectively combine structural information and subcellular distribution with molecular mapping are those of Albert et al.44,45, which looked at the structural heterogeneity of the proteasome in algae relative to its subcellular location. They identified three distinct subpopulations based on their subcellular location: proteasomes tethered to the nuclear envelope, proteasomes tethered to the basket of the nuclear pore complex and proteasomes localized in clusters near the endoplasmic reticulum. Those of the last group are more often in the active state and contain additional densities, which the authors conclude are protein substrates from the endoplasmic reticulum that are degraded by the proteasomes45. In a similar study, Guo et al.46 identified a local increase of proteasomes around poly-GA aggregates in primary rat neurons, which are more often in a stalled conformation.

Cellular ultrastructure

Besides structural information on macromolecules, cryo-ET on FIB-fabricated lamellae provides data on cellular morphology at nanometer resolution and allows for small ultrastructural features to be studied in a near-native state at a resolution currently not yet feasible with other cellular imaging techniques such as serial FIB scanning electron microscopy (SEM) imaging47. An example is the recent study on the effect of several cortical endoplasmic reticulum–plasma membrane tethers on endoplasmic reticulum membrane morphology48. The authors identified that the proteins tricalbin 1–3 are responsible for forming sharp cortical endoplasmic reticulum membrane curvatures of ~15 nm facing the plasma membrane. These areas were found to be involved in protecting the cell against heat shock, presumably by allowing transfer of lipids from the cortical endoplasmic reticulum to the plasma membrane. The morphology of the endoplasmic reticulum–plasma membrane contact sites formed by these tethers is difficult to study with classical electron microscopy methods, which can change the membrane morphology. Khanna et al.49 observed nanometer-sized morphological changes in the peptidoglycan layer of Bacillus subtilis during early sporulation: by blocking peptidoglycan synthesis with antibiotics during sporulation, they found that the peptidoglycan layer was not fully degraded at the start of engulfment.

Other examples of cryo-ET on FIB-fabricated lamellae to study cellular morphology include the type 6 secretion system of Amoebophilus asiaticus in phagosomal escape in infected amoeba cells50 and polyglutamine inclusions associated with Huntington’s disease in the morphology of mammalian and yeast cells51,52. Cryo-ET of FIB-fabricated lamellae was also used to study the morphology of a phage replication compartment in infected Pseudomonas chlororaphis and to show how viral capsids migrate from the bacterial membrane toward this compartment via virally encoded, tubulin-like filaments53,54.

In summary, cryo-ET on FIB-fabricated lamellae enables structural and ultrastructural studies of the cell interior in a near-native state. These studies were enabled by the development and incremental improvement of lamella fabrication, data processing and subtomogram averaging, all of which are discussed in the next section.

Recent advances in cryo-ET on FIB-fabricated lamellae

Obtaining a cellular subtomogram average structure of a macromolecule from a biological sample requires several steps, each of which must be reliable for success. For low-copy-number macromolecules, a large number of tomograms need to be acquired, underlining the importance of sample throughput. In this section, we describe the latest improvements in sample preparation (specifically vitrification and lamella fabrication), tilt-series acquisition and data processing and how to find macromolecules in the tomographic volume of the crowded cellular environment.

Sample vitrification

For cryo-ET on FIB-fabricated lamellae, cells need to be applied to or grown on an electron microscopy grid compatible with a transmission electron microscope and FIB/SEM. Gold or titanium55 grids are typically used because they are nontoxic to cells, stronger and less expensive. Only single cells, or small clusters of cells, that are located near the center of the grid and center of the grid squares are suitable for FIB milling, which can greatly reduce the number of sites available for lamella fabrication15. Photo-micropatterning can be used to make cells selectively adhere to the center of the grid squares55.

Cells on an electron microscopy grid are then rapidly frozen to preserve the fine structural and ultrastructural features. Ideally, all cells end up in a vitreous state, which preserves them under near-native conditions without the formation of crystalline ice, which would disrupt the structure of the cell4. The most commonly used ultrarapid freezing technique for cellular samples is plunge freezing, which involves quickly submerging the sample into liquid ethane or a mixture of liquid ethane and propane and cooling with liquid nitrogen to cryogenic temperatures56,57,58. A limitation of plunge freezing is that it does not have a cooling rate high enough to fully vitrify samples more than a few µm thick, such as most mammalian cells20. The use of a cryoprotectant such as glycerol can benefit the vitrification of cells with plunge freezing59. Higher cooling rates have been observed with jet freezing, where a cryogen is jetted onto the center of the sample34,60,61. Optimization of jet freezing or other ultrarapid freezing techniques such as slam freezing62,63 may potentially improve the vitrification of eukaryotic cells for cryo-ET.

High-pressure freezing (HPF) can vitrify thicker samples, such as multicellular organisms, tissue and organoids, by greatly increasing the pressure to 210 MPa less than 1 s before ultrarapid freezing of the sample, which inhibits expansion of the volume during crystalline ice formation, effectively lowering the cooling rate that is required to form vitreous ice64. However, HPF generally results in a thick layer of ice surrounding the biological sample that can be too large to remove with the FIB. Harapin et al.65 describe how a biological sample high-pressure frozen in 2-methylpentane can be thinned by sublimation of the 2-methylpenthane in a vacuum at temperatures below the devitrification temperature of ice. However, the biological sample is exposed to an organic solvent during the freezing process and requires long milling times (>1 d) for bulkier samples with a conventional gallium FIB. HPF can be used to ultrarapidly freeze smaller cells with less ice surrounding them to allow direct lamella preparation by sandwiching the grid between two polished planchettes66. Another strategy involves cutting back the planchette with a cryo-ultramicrotome so that lamellae can be prepared near the cutting edge67.

FIB fabrication of lamellae

After freezing, lamella fabrication can be performed with a FIB–SEM dual-beam microscope. The FIB is used to remove all of the material above and below part of the cell, leaving a ~200 nm layer of vitreous cellular material suitable for cryo-ET14,15 (Fig. 1e–h). The lamella surface area depends on the cell type, the chosen lamella width and the milling angle. For mammalian cells, the width is typically between 8 and 20 µm and the length is between 5 and 40 µm. Because the commonly used gallium FIB has a Gaussian distribution68, lamellae are often thinner at the front compared with the back. A more even lamella thickness can be achieved by slightly tilting the sample during the milling process to produce a wedge geometry, followed by a fine milling step at the original milling angle14. Another common problem is the occurrence of cracks in lamellae caused by a build up of tension in the grid foil. This cracking can be mitigated by creating micro-expansion joints during the first milling step to release the tension69 (Fig. 1g,h).

Manual FIB fabrication of lamellae is a time-consuming process. An experienced user can create 5–16 lamellae per day69,70. Automated lamella fabrication71,72,73 can run overnight and greatly increase throughput. However, the number of lamellae milled from a single grid is limited by the rate at which water vapor is deposited on the sample due to the high partial H2O pressure conditions in the sample chamber. Cryo-FIB–SEM microscopes for lamella fabrication can have a continuous deposition of water vapor on the finished lamellae at a rate of up to 85 nm h−1 (refs. 14,73). The accumulation of the ice layer, as well as deposition of sputtered material on the lamella, can be decreased by performing the final milling steps of each lamella last, but it still in practice limits the number of lamellae from a single grid to 3–5.

Bringing the sublimation and deposition rates in the FIB–SEM chamber closer to equilibrium can reduce the ice growth rate. Tacke et al.73 accomplished this by installing additional cryo-shields and a cryo-shutter (which reduces the opening for the ion beam) in the FIB–SEM sample chamber to decrease the partial water vapor pressure and by using a heater to heat the sample to −165 °C; these modifications decreased the ice growth rate to 5.6 nm h−1. The ice growth rate could be further reduced to 0–2 nm h−1 by employing a different cryo-shutter that further decreases the opening for the ion beam and completely blocks the electron beam. Additionally, the authors showed reduced ice contamination by using a glove box flushed with nitrogen gas for sample handling and by performing sample unloading in a vacuum during the brief transfer from the transfer rod to the liquid nitrogen in the loading station73. Ice growth rates of <2 nm have also been reported recently35.

Lamellae of larger samples can be created by extracting biological material with a FIB lift-out approach. With this method, a 3- to 5-µm-thick lamella is created perpendicular to the surface of a high-pressure frozen sample and transferred to a slot on another grid with a cryo-micromanipulator74,75,76,77 or cryo-gripper78,79, followed by fine FIB fabrication of lamellae. The thick ice layer can be removed with a cryo-microtome or freeze fracturing before the lift-out procedure80,81. Since high-pressure-frozen samples are generally surrounded by ice, the biological specimen cannot always be found with FIB or SEM imaging, resulting in a substantial risk of missing the cells entirely. CLEM can be used to determine where fluorescently labeled cells are located in the ice71,75,78.

While gallium ion beams are most commonly used for cryo-FIB fabrication of lamellae, a plasma FIB using other elements (for example, xenon, argon, nitrogen or oxygen) may speed up lamellae fabrication and bulk removal for lift-out approaches78. Xenon plasma FIBs are reported to have faster milling rates compared with gallium, with a decreased depth of the damage layer in material science applications82,83. A recent study showed that argon is suitable for preparing FIB-fabricated lamellae for subtomogram averaging in an automated fashion, enabling a 4.9 Å structure of the 80S human ribosome in HeLa cells to be determined35.

Cryo-ET on FIB-fabricated lamellae

After lamella fabrication, tilt series of the lamellae are acquired with a cryogenic transmission electron microscope. Cryo-ET benefits from the same technical improvements that enabled the resolution revolution in single-particle cryo-EM, in particular direct electron detectors and processing software (see the section ‘Data processing for subtomogram averaging’)1,84. A limitation of tomography is the restricted tilt range of ~120°, which results in missing information in reciprocal space in the direction of the electron beam (that is, a missing wedge). This missing wedge results in lower resolution in the direction of the electron beam and reconstruction artifacts in tomograms. Limitations of the missing wedge can be alleviated for randomly oriented macromolecular complexes (but not for those of the preferred orientation) during subtomogram averaging.

The use of an energy filter can increase the contrast in thicker samples by removing inelastic scattered electrons, which is particularly beneficial for cellular cryo-ET. A phase plate, which (ideally) shifts electrons 90° in phase to maintaining a good transfer of low-spatial-frequency information, can be used to acquire tilt series close to focus while retaining good contrast. For the most commonly used Volta phase plates, this is achieved by creating a beam-induced Volta potential on a heated carbon film of ~12 nm in the back focal plane of the transmission electron microscope85. The disadvantages of Volta phase plates include the scattering of electrons from passing through the carbon film, the technical challenges of maintaining a consistent focus and phase shift due to variability in the phase shift from charging of the carbon film and the increased difficulty of accurate contrast transfer function (CTF) estimation and correction of the data86. A 5–20 nm platinum layer applied on lamellae after FIB milling improved the conductivity and decreased sample charging during Volta phase plate experiments14. Early results with laser phase plates, which use the standing light wave from a laser rather than a film to create a phase shift, indicate that they have the potential for more stable phase shifts, without electron scattering from a carbon film87.

The transfer of high-resolution information in the tilt series can be improved by adopting a dose-symmetric tilt scheme where the lowest tilt angles are imaged first. This strategy minimizes the cumulative damage from the electron beam at the angles where the sample is effectively the thinnest because of the geometry of the lamella86,88. Dramatic increases in the rate of tilt-series acquisition to several minutes per tilt series have been reported with fast-incremental tilt schemes, where autofocus and tracking steps are omitted and movie frames are continuously recorded while stopping at different tilt angles to record an image by unblanking the electron beam89,90. The use of this tilt scheme enabled a subtomogram-averaged structure of isolated ribosomes at subnanometer resolution on a dataset of 12 tilt series recorded in 50 min90. This rapid acquisition time contrasts with commonly used tilt schemes requiring 20–60 min per tilt series. Another alternative faster tilt scheme uses beam shift (with correction for variable beam tilt as a function of beam shift) to acquire several images per tilt, effectively allowing tilt series to be acquired in 5 min without sacrificing data quality91. Montage tomography using beam shift, optimized to decrease fringe artifacts on the electron beam edges and with minimal excessive electron dose on areas exposed multiple times, is also being developed to acquire volumetric data over a larger area on FIB-fabricated lamellae92.

Data processing for subtomogram averaging

Data processing aims to correct some of the aberrations caused by the electron microscope during imaging to improve the resolution of the data and the resulting structures. Irradiation by the electron beam introduces motions to the sample during the acquisition. A series of images acquired at a single tilt angle with a direct electron detector can correct for this beam-induced motion in a process termed motion correction93. Electrons also cause cumulative damage to the sample during the acquisition of a tilt series, which can be compensated for with dose weighting94, where the more dose-sensitive, high-spatial-frequency information of fine molecular details is downweighed in tilts acquired later. Finally, image acquisition with an electron microscope results in an oscillating pattern of contrast loss and reversals as a function of spatial frequency (CTF), which is strongly affected by the defocus8. CTF correction is essential to achieve high resolution in a subtomogram-averaged structure. Where local variations in defocus from sample tilt are corrected for by most methods95,96, more advanced 3D CTF correction methods can better adjust for the local defocus of a particle by taking into account its height (z) within the sample97,98,99,100,101.

Tilt series also need to be accurately aligned for 3D tomogram reconstruction, since this affects the final resolution86,99,100. Accurate local alignments are preferred over global alignments to better correct for complex sample motions and deformation during tilt-series acquisition101. Tilt-series alignment is most commonly performed with patch tracking, which tracks cellular features based on cross-correlations15,96. However, more accurate local alignment can be achieved with fiducial-based alignment. Colloidal nanogold particles (5–20 nm) act as fiducials when mixed with small cells (such as bacteria) before ultrarapid freezing, but do not give a good distribution for larger eukaryotic cells. Endocytosis of bovine serum albumin–nanogold by mammalian cells before ultrarapid freezing can be used to distribute fiducials throughout the cell, which can be used for tilt-series alignment after lamella preparation102. To apply gold fiducials to lamellae after fabrication, nanogold can be resuspended in 2-methylpentane and applied to the surface of the lamellae65.

Numerous software packages have been developed over the years to obtain structures of macromolecules from tomograms94,99,100,101,103. The recently developed software packages EMAN2, emClarity and M correct imaging artifacts locally, per particle99,100,101, which improves the resolution that can be obtained with subtomogram averaging. For example, the software package M refines several per-particle parameters, including alignment and CTF correction, in an iterative process and constrains these based on other particles from the same tilt series101. Artifacts from intermediate interpolations are minimized by directly reconstructing the final subtomogram average map from the motion-corrected tilt series rather than extracted subtomograms. When several different local corrections are applied together, this combination results in a sizable increase in resolution99,100,101. Promising subnanometer-resolution structures have also been obtained with tomography and single-particle hybrid approaches104,105, but their use on lamellae has not yet been reported.

Interpreting tomograms of the crowded cellular environment

While visualizing the native cellular environment is the main strength of cryo-ET on lamellae, it also makes interpretation of the crowded cellular environment challenging. Numerous different macromolecules in the cell and the typically low signal-to-noise ratio in cellular tomograms make it difficult to find specific macromolecules. Localizing macromolecules (often done by hand) is essential for subtomogram averaging. The identifiable macromolecules depend on many factors, including size, shape, the relative density compared with the surrounding medium, the local crowdedness of the tomogram and experience and biological understanding of the researcher.

A proven method for identifying the location of macromolecules in room temperature electron microscopy is the use of nanogold particles conjugated to antibodies that specifically bind to the macromolecule in a cellular section (that is, immunogold labeling). Robust labeling strategies for cryogenic applications may be of considerable benefit in localizing and identifying macromolecular complexes. It has been demonstrated that SNAP-tag-labeled nanogold106 and immunogold labeling on live cells before freezing can label proteins on extracellular membrane surfaces107,108. A recent study also showed that DNA origami structures with a high affinity for green fluorescent protein (GFP) can be used to specifically label proteins on the surface of cells and viruses before sample freezing109. Iron-binding ferritin fusion proteins have also been shown to enable intracellular labeling of bacteria visible in cryo-EM110. In principle, a fluorescent label combined with high-accuracy cryo-CLEM can be used to find macromolecules of interest (discussed in the section ‘CLEM’).

In addition to labeling, unknown proteins in a complex have been identified in lamellae using genetic modification. For example, deletion mutants enabled the identification of FraD as a structural component of the plug of the septal gap junction protein in cyanobacteria39. In another study, expression of an nsp3-EGFP fusion protein40 facilitated the discovery of nsp3 as a structural component of the pore that spans the coronavirus replication compartment.

Identifying macromolecules of interest can be facilitated by various computational methods (reviewed in ref. 111). One such approach denoises tomograms with filters to greatly increase their visual clarity and improve the signal-to-noise ratio96,100,112,113. More recently, image restoration based on machine learning has shown promising results114,115,116. Computational subtraction can diminish artifacts resulting from the thin platinum layer sputtered on the finished lamellae14, contaminants and/or nanogold fiducials96,117.

Finding known macromolecular complexes can be automated with template matching, where the density of a known macromolecule is compared with the tomographic volume at different orientations by means of cross-correlation103. The chance of success depends on different factors, including the macromolecule (for example, size, shape, the distinctness of the density and the subcellular environment) and the tomographic data (for example, quality, resolution and crowdedness). Algorithms that use machine learning118,119,120, pattern mining121 and software for identifying membrane-bound and transmembrane macromolecular complexes122 have been developed recently to assist in finding unknown macromolecules without an existing template.

Another process that can assist in interpreting the cell environment is segmentation, where cell membranes are traced in the volume, to enable quantifications of membrane curvature, surface area and distance of macromolecules to membranes. While manual segmentation is time consuming, (semi-)automated methods have been demonstrated123,124 and successfully applied in different studies48,49,51,54. The quality of segmentation can be improved with post-processing, such as low-pass filtering or mean curvature motion125.


Cryo-EM methods can achieve atomic-resolution structures, but the low signal-to-noise ratio can make it difficult to locate macromolecules or biological structures. Likewise, both FIB and SEM imaging are limited to obtaining morphological information of the cell surface. Cryo-CLEM can complement these imaging modalities by localizing fluorescently labeled macromolecules and/or biological structures in the sample and providing contextual information about the cell interior. CLEM techniques were first developed for classical electron microscopy on chemically or cryogenically fixed, resin-embedded samples at room or low temperature followed by room temperature electron microscopy observation126, and later adapted for use under cryogenic conditions127,128. This adaptation came with a unique set of challenges, in particular for super-resolution light microscopy. In this section, we first describe different applications for CLEM in the context of cryo-ET on FIB-fabricated lamellae, followed by the challenges and developments in pushing the resolution with cryogenic light microscopy (cryo-LM).

CLEM applications for cryo-ET on FIB-fabricated lamellae

One application of CLEM for FIB-fabricated lamellae is to use light microscopy to help determine which cells to use for FIB fabrication of lamellae, for which a coarse correlation accuracy (up to several micrometers) between light microscopy and SEM data is sufficient. Examples include finding the subcellular localization of actin waves129, different disease-associated fluorescently labeled mutants of proteins51,52, cells in a specific stage of the cell cycle130 or cells infected with a pathogen (Fig. 3a,b).

Fig. 3: Three examples of distinct applications of CLEM for cryo-ET on FIB-fabricated lamellae.
figure 3

White dashed boxes indicate the sections that are magnified in the next panel, with a solid white line indicating the bottom side of the next panel. Note that the enlarged area in f from e has been rotated. a,b, 2D–2D correlation of cryo-LM with a SEM image to find cells of interest before FIB milling. Macrophages were differentiated from primary blood monocytes grown on an electron microscopy grid and infected with Mycobacterium marinum. The nuclei are labeled in blue and the bacteria are labeled in green. Scale bars, 300 µm (a) and 50 µm (b). c, Targeted FIB milling approach using CLEM to make a lamella of a cell containing a macromolecular complex or structure of interest. A cryo-LM z stack is correlated with a FIB image at the milling angle using fiducials visible with both the FIB and light microscopy (red circles) and projected into a 2D image. This 2D projection is overlaid with a FIB image and can be used to target the position for a FIB-fabricated lamella so that it contains a biological target of interest. In this example, macrophages were infected with Escherichia coli expressing a fluorescent protein. Scale bar, 10 µm. d, FIB image after FIB fabrication of a lamella overlaid with the projected 2D fluorescence image via the image in c using the software eC-CLEM155. Triangular markers were prepared with FIB (white arrows) on the side of the lamella at the start of fabrication to determine where the final lamella should be made in subsequent finer milling steps. The presence of fluorescence on the final lamellae suggests that the targeted fluorescent bacteria are probably contained within the lamella. This was later confirmed to be the case with cryo-EM (data not shown). Scale bar, 5 µm. e, 2D–2D correlation to find areas for recording tilt series containing a macromolecule, complex or structure of interest. In this example, the aim was to find preperoxisomal vesicles by overexpressing the peroxisomal membrane protein-14 (Pex14) fused to GFP (green) in yeast (Hansenula polymorpha ΔPEX19). The cell wall was labeled in blue to correlate cryo-LM to electron microscopy images. Cryo-LM was acquired after FIB fabrication of a lamella so that only fluorescence from macromolecules still present in the lamella was retained. A 2D cryo-LM image was correlated to the SEM overview of the lamella after final fabrication using the software eC-CLEM155. Scale bar, 5 µm. f, Low-dose cryo-TEM overview of the same lamella (enlarged from the boxed area in e) correlated to the cryo-LM image to determine where to record the tilt series. Scale bar, 1 µm. g, Summed tomographic slices from a tilt series recorded on the lamella (enlarged from the boxed area in f) correlated to the green fluorescent signal to indicate where preperoxisomal vesicles should be located within the tomogram. Both the nucleus (Nu) and a mitochondrion (Mi) are clearly visible. Scale bar, 500 nm. h, Enlarged section of the same summed tomographic slices (dashed box in g) where several preperoxisomal vesicles are visible (white arrows). A lipid droplet (LD) is surrounded by the endoplasmic reticulum and the cell is surrounded by the cell wall (CW). Scale bar, 100 nm.

Light microscopy data can be acquired before ultrarapid freezing51, which allows the use of oil immersion objectives with a high numerical aperture, as well as live imaging over time26. The major disadvantage is that the live sample changes over time between light microscopy acquisition and sample freezing, making the correlation inaccurate for fast and dynamic processes. The time between imaging and freezing can be minimized to seconds using a plunge freezer with an integrated fluorescence microscope131 or to milliseconds using a cryo-fixation microfluidics device that freezes the sample with a liquid nitrogen-cooled heat sink79. The cooling rate of this microfluidics device was sufficient to vitrify larvae of C. elegans in 10% glycerol, which could subsequently be imaged with cryo-ET with a FIB lift-out approach.

Light microscopy data can also be acquired post-vitrification, where all biological processes are arrested. To enable this, dedicated cryogenic stages127,128,132,133,134 maintain the sample below the devitrification temperature, minimize drift, allow for fast cryogenic transfers and provide a low-humidity environment to minimize ice contamination on the sample while imaging. Integrated approaches, where light microscopy is combined in a transmission electron microscope or FIB–SEM microscope, can also minimize ice contamination by decreasing cryogenic transfer and allow cryo-LM imaging in a vacuum environment, enabling easier and faster correlations81,135.

Since lamellae have a local thickness of 50–300 nm and cells are typically micrometer sized, only a fraction of the total cellular volume is contained in a prepared lamella. Macromolecules, cellular structures and interactions of interest that are not ubiquitously expressed throughout the cell may have a low chance of being contained within the prepared lamella. Cryo-CLEM can help pinpoint where to prepare lamellae to contain the target30,46. One approach uses electron-dense, micrometer-sized fluorescent fiducials visible in both the FIB–SEM and light microscopy to calculate the transformation parameters (scaling, translations and rotations) to correlate 3D light microscopy data to a 2D FIB image (Fig. 3c,d)134. Arnold et al.134 reported a ≥150 nm localization accuracy on 60% of targets (polystyrene beads) with this method. The accuracy on intracellular targets may be lower because of optical distortions from refractive index mismatches between the cell and the gaseous environment136 and sample distortions during FIB fabrication of lamellae71. This approach requires axial localization of fiducials and the structure of interest with an accuracy of hundreds of nanometers to achieve consistent targeting, depending on the size of the target. A recent study reported a ~91% success rate in targeting the centrosome for FIB fabrication of lamellae, by using a widefield microscope mounted underneath the sample in the FIB–SEM microscope, which can be operated during milling135. Targeting is achieved by first calibrating the coincidence point for the FIB, SEM and light microscopy, followed by focusing and centering of the intracellular target of interest135,137. Serial cryo-FIB–SEM imaging, where thin layers of cellular material are removed with the FIB, followed by SEM imaging to obtain 3D data, has also been successfully employed to help determine where the final lamella is formed138. More importantly, it provides morphological context of the cellular material that is otherwise removed without imaging, and can be correlated with cryo-LM to obtain specific information about the cell.

In another application, cryo-CLEM is used to identify areas of interest in the prepared lamellae, which can help to determine where to acquire tilt series or to identify macromolecules in the crowded cellular environment (Fig. 3e–h). Only a low correlation accuracy below the field of view used for tomography (typically 1–2 μm) is required for its successful application. A CLEM approach combined with split-Venus labeling was used to identify areas where nuclear core complexes interact with autophagosomes in the final lamellae to record tilt series of it30. Current challenges for acquiring fluorescence data after FIB fabrication of lamellae are that the fluorescent signal may be difficult to detect and that cryogenic transfers to a cryo-LM may directly contaminate the lamellae with ice42,81. In another approach where fluorescence light microscopy data are acquired before FIB fabrication of lamellae, a tilt-corrected optical slice is extracted that matches with the final lamella. The tilt of the lamella is determined by recording transmitted light brightfield data of the lamella after transmission electron microscopy (TEM) imaging42. The downside of recording fluorescence data before lamella fabrication is that the signal from out-of-focus objects that are not present in the lamella may still be present in the correlated optical slice due to the limited axial (z) resolution of light microscopy42. In principle, cryo-LM can be correlated to tomographic data to help pinpoint the 3D location of labeled macromolecules within tomograms. In practice, this is not yet feasible due to the limited resolution of cryogenic fluorescence microscopy.

Pushing the resolution beyond the diffraction limit in cryo-LM

Since the correlation accuracy of CLEM applications for FIB-fabricated lamellae depends greatly on the resolution of the light miscroscopy, increasing the resolution obtained under cryogenic conditions is critical. In light miscroscopy at room temperature, the diffraction-limited lateral resolution is normally 200–300 nm, which is largely dependent on the wavelength of the emitted fluorescence and the numerical aperture of the objective. The resolution obtained for cryo-LM is typically lower due to several challenges at cryogenic conditions.

One limitation is the difficulty of developing immersion objectives for cryogenic conditions with a high numerical aperture, which limits the photon yield and resolution. Where oil immersion objectives with a numerical aperture of ~1.4 can be employed at room temperature, only air objectives, for which the numerical aperture is fundamentally limited to below 1, can be used for cryo-LM. Several approaches have been explored to overcome this, such as the use of an immersion objective together with cryogenic immersion fluids with a matching refractive index139,140 or an air objective together with a superhemispherical solid immersion lens to bridge the distance between the objective and the sample141.

Imaging beyond the diffraction limit at room temperature was enabled by the development of several super-resolution techniques142,143,144, many of which have been successfully demonstrated at cryogenic temperatures. Although these super-resolution techniques have not yet been reported for cryo-ET on FIB-fabricated lamellae, they can benefit accurate CLEM (reviewed in ref. 136). The application of super-resolution light microscopy in CLEM approaches for FIB-fabricated lamellae can provide a substantial increase in resolution and localization precision and help to reduce the current resolution gap with cryo-EM. Considerable progress has been made in the development of localization-based super-resolution cryo-LM139,141,145 and used to study biological samples, including the type 6 secretion system in Myxococcus xanthus and Vibrio cholerae146, localization of TOM20 in cryogenic sections of mammalian cells147 and several bacterial proteins in Caulobacter crescentus148. A recent study showed successful cryo-CLEM on tomograms recorded on the mammalian cell periphery with a localization precision of 30 nm149. One challenge for localization-based super-resolution cryo-LM is that fluorophores have altered photoswitching behavior compared with at room temperature, which requires the use of high laser powers to make them adopt a dark state139,140,145,146,147,149 and thus may devitrify the sample146,147,149. One study reported that photoswitchable fluorescent proteins can be forced into their deactivated state and retained for at least 7 d after freezing by illuminating the sample before ultrarapid freezing149. Super-resolution optical fluctuation imaging, which pinpoints fluorophores based on bulk fluorescence intensity fluctuations rather than isolated blinking events, has been applied to lower the laser power as well as acquisition times. Super-resolution optical fluctuation imaging has been successfully demonstrated under cryogenic conditions to visualize conventional cryo-EM samples without cryoprotectant, with a resolution up to 135 nm150. The nonlocalization-based super-resolution techniques structured illumination microscopy and Airyscan confocal microscopy have also been successfully applied under cryogenic conditions, resulting in lateral resolutions of up to 200 and 290 nm, respectively, using air objectives151,152.

The future of cellular cryo-ET

While cryo-ET on FIB-fabricated lamellae has developed into a powerful technique over the past few years to elucidate the profound structural complexity of the cell, its full potential has not yet been achieved. The resolutions obtained with subtomogram averaging are steadily improving35,36,38,101 and subnanometer or even near-atomic resolution in situ structures for many macromolecules can be attained using cryo-ET on FIB-fabricated lamellae. One of its current limitations is low throughput due to the limited speed and success rate of each individual step.

First, sample preparation of thick cells, organoids, multicellular organisms and tissues requires further improvement. Incomplete vitrification of large cells with plunge freezing limits the throughput and cell types that can be effectively studied. Adaptation of HPF sample holders66 or the development of other ultrarapid freezing techniques61,62,79 compatible with subsequent cryo-ET provides hope for improved vitrification of more (and larger) cell types. While impressive results have been obtained with FIB lift-out approaches on multicellular organisms, the current success rate of ~20% is too low for most in situ structural studies78.

To increase throughput, the rate of FIB fabrication of lamellae could be increased with an automated FIB–SEM dual-beam microscope71,73 with minimal ice growth and contamination from transfer73. Faster tilt-series acquisition schemes89,90,91 can substantially speed up tomography data acquisition.

Further automation of data processing, tomogram reconstruction, data analysis and subtomogram averaging, which are arguably larger bottlenecks than data acquisition, would also improve throughput. The application of machine learning for image restoration and denoising114,115,116, map denoising101, locating macromolecules in tomograms119,120,121,122 and automated reconstruction153 could help improve the speed and accuracy of the different processing steps. The rapid improvements for machine learning applications, such as the recent success of AlphaFold 2 (ref. 154) in predicting protein structures, will also facilitate these advancements.

With increased throughput and structures of increased resolution and from a greater range of macromolecules, new types of research will become feasible. For example, in situ structure determination of smaller macromolecules and the detection of more subtle conformational states may be possible. Hundreds or even thousands of different macromolecules (both known and unknown) in different conformational states could be identified in tomographic volumes and linked to their subcellular localization, bringing visual proteomics closer to becoming a reality43. This in situ structural information could be established over a wide range of species and cell types, for different diseases and under different physiological conditions and stimuli. Effective sharing and open access to annotated cellular tomographic data could accelerate such an endeavor101.

In summary, developments in cryo-ET and FIB fabrication of lamellae over the past years have demonstrated remarkable advancements, enabling in situ structural studies that were not previously possible. Major challenges are still present, but resolving the current shortcomings can greatly expand what is possible in in situ structural biology and advance our understanding of macromolecular structures in relation to their function and subcellular localization—effectively bridging the gap between structural and cell biology. Cryo-ET on FIB-fabricated lamellae, and its contributions to in situ structural biology, may soon prove to be the second wave in the cryo-EM revolution.