Review | A TRENDS Guide to Imaging Technologies Trends in Biotechnology Vol. 20 No. 8 (Suppl.), 2002 In vivo veritas: electron cryotomography of cells Jürgen M. Plitzko, Achilleas S. Frangakis, Stephan Nickell, Friedrich Förster, Ariane Gross and Wolfgang Baumeister

Automated electron cryotomography enables the study of organelles or whole cells embedded in vitrified ice in a close-to-life state and with a resolution of a few nanometers. This technology will provide new vistas of the supramolecular architecture inside cells that orchestrates higher cellular functions.

The long prevailing view of a as a membrane-bound ‘weighted backprojection’ algorithm [3] – the projections reaction compartment filled with freely diffusing and col- must be mutually aligned to establish a common system liding macromolecules can no longer be maintained [1]. of coordinates (Fig. 1). There is growing awareness that fundamental cellular The relationship between resolution d of a tomographic functions are carried out by ensembles of macromolecules, reconstruction of a cylindrical object of diameter D, from N protein complexes or ‘molecular machines’. And, as in a equally spaced projections covering the full angular range of factory, the operation of these machines must be coordi- 180º, is given by the Crowther relationship d ~π D/N [4]. nated to give rise to a stochastically variable supramolecular For the reconstruction of an object of 200 nm diameter architecture. On this level of structure, the cell is, by and with a resolution of 2 nm one would therefore need 320 large, an uncharted territory. None of the existing imaging projections and tilt increments of 0.6º. For an optimal techniques enables the study of pleiomorphic structures, sampling of objects of constant thickness, non-equidistant such as organelles or whole cells, with a resolution of a tilt angles are advantageous [5]. Owing to physical con- few nanometers, as is required for identifying macromol- straints (e.g. limitations of specimen holders and increase ecules in situ and for describing their interaction networks. of specimen thickness at higher tilt angles) the full angular It has become clear from recent proteomics studies [2] that range cannot be accessed; in practice, it is restricted to we have vastly underestimated the extent to which indi- about ±70°. Considering the projection theorem [6], the vidual proteins assemble to form distinct complexes. The limited tilt range implies that data are missing in a methods traditionally used in biochemistry for isolation and wedge-shaped region of the Fourier space resulting in purification of protein complexes tend to select for the distortions of the reconstruction, most notably a loss of stable and the abundant ones; the rare or transiently asso- resolution in the z-direction (discussed later). ciated complexes or the ones held together by forces too A key problem in ET, which for more than two weak to withstand the procedures escape detection.There- decades stood in the way of establishing it as a practicable fore, there is a strong incentive to develop methods – ideally technique, is reconciling two requirements that are in con- non-invasive – to study supramolecular architecture in a flict with one another.To obtain ‘high’ (<5 nm) resolution cellular context. tomograms with minimal distortions, one has to collect Jürgen M. Plitzko, data over as wide a tilt range as possible and with incre- Achilleas S. Frangakis, Electron : principles and problems ments as small as possible.At the same time, the cumulative Stephan Nickell, Given the large depth of focus, electron microscopic images electron dose must be kept within tolerable limits to prevent Friedrich Förster, are essentially 2D projections of the object in the direction radiation damage erasing the finer details of the structure Ariane Gross and of the electron beam. Features from different levels within or, in the worst case, rendering reconstructions meaning- Wolfgang Baumeister* the object overlap and cannot be separated. Tomographic less. Samples embedded in amorphous ice are extremely Dept of Molecular Structural techniques acquire projections of the object as viewed from sensitive to radiation damage and should not be exposed Biology, different directions and then merge them computationally to >2000–5000 e nm−2 (e = electrons) for visualizing Max-Planck-Institute of to obtain a 3D reconstruction of the object volume. In molecular structures. Biochemistry, D-82152 (ET) the specimen holder is tilted The successful realization of electron cryotomography Martinsried, Germany. incrementally around an axis perpendicular to the electron is based on dose fractionation. Early theoretical consider- *e-mail: beam and images are recorded at each position. Before the ations [7] had suggested that, in principle, the electron [email protected] 3D density map can be calculated – most commonly by a dose, which is required to visualize structural features

S40 0167-7799/02/$ – see front matter ©2002 Elsevier Science Ltd. All rights reserved. PII: S0167-7799(02)02017-6 http://www.trends.com Trends in Biotechnology Vol. 20 No. 8 (Suppl.), 2002 A TRENDS Guide to Imaging Technologies | Review

close to the resolution limit, is the same for 2D and 3D images containing equivalent information.Thus, the separ- (a) (b) ation of features in z-direction would be a bonus one has not to pay for! By combining the information of all pro- jections of a tilt series, the signal-to-noise ratio improves w w as it does by adding up statistically noisy 2D images. More w w w w recent computer simulations have confirmed that the dose w w w fractionation theorem holds, even in the case of (not too strongly) absorbing specimens [8]. Therefore, one could, in principle, fractionate the dose over as many projections as an optimized tilt geometry might require. There is, TRENDS in Biotechnology however, a practical limitation: the signal-to-noise ratio must still enable alignment of the projections accurately enough to establish a common system of coordinates. Vitrification by rapid freezing ensures close-to-life preser- Figure 1. Principle of This is done by cross-correlation, and facilitated by vation of the structures to be studied and enables the electron tomography The figure shows a single axis adding fiducial (gold) markers to the specimen. capture of dynamic events [18]. It avoids the risk of arti- tilt series data acquisition The alignment problem is aggravated by the limited facts traditionally associated with chemical fixation and scheme. The object is mechanical accuracy of specimen holders. Deviations from staining or with the dehydration of cellular structures. represented by a pleiomorphic object (represented by a DNA an ideally eucentric tilt geometry,instability and drift of the Equally important, tomograms of frozen-hydrated struc- knot) to emphasize that electron cryoholders cause movements in the x-, y- and z-direction tures represent their natural density distribution while tomography can retrieve 3D information from non-repetitive that translate into lateral shifts and changes of focus.There- staining reactions tend to produce intricate mixtures of structures. (a) A set of 2D fore, following each change of tilt angle, the specimen positive and negative staining, and therefore compromise projection images is recorded while the specimen is tilted (or the image) has to be realigned and refocused.This can interpretation in molecular terms. incrementally around an axis be done manually,but at the expense of additional exposure A major limitation in electron tomography is specimen perpendicular to the electron beam, and projection images of of the specimen to the electron beam. In fact, the dose thickness and here one has to take into account that for the same area are recorded on ‘overhead’ spent on these alignments can far exceed the simple geometric reasons a specimen with a thickness of a CCD camera at each tilt ± ° dose spent usefully on recording data. 200 nm (at 0º tilt) will increase in thickness to 540 nm angle. Tilt range is typically 70 and the tilt increments 1.5–3.0°. at 70º tilt. A thickness larger than the mean free path of The backprojection method Instrumentation and automation electrons brings us into a multiple scattering regime. In explains the principle of 3D reconstruction in a fairly intuitive With the advent of computer-controlled transmission particular, the contribution from inelastically scattered manner. (b) For each weighted electron microscopes equipped with eucentric goniometers, electrons (the inelastic mean free-path is ~200 nm at an projection (W), a backprojection body is calculated, and the sum and the availability of large-area charge-coupled device acceleration voltage of 120 kV; for 300 kV it increases to of all projection bodies (CCD) cameras, it became possible to develop and imple- 350 nm) [19] can degrade the quality of the images and represents the density distribution of the original object ment complex image-acquisition procedures running in a tomograms quite significantly. Here, energy filters offer – the tomogram. fully automated fashion [9,10]. This made possible the some remedy: when operated in a zero-loss mode, the ad- recording of tomographic datasets with the specimen re- verse contribution (‘blurring’) of the inelastically scattered maining centered and at a uniform level of focus, while electrons can be removed and contrast is improved. Samples the cumulative electron dose is kept within tolerable much thicker than 0.5–1.0 µm cannot be studied in toto limits [11]. At lower magnifications suitable for imaging and, therefore, must be cryosectioned before tomographic large structures at the resolution of ~5 nm, datasets com- analysis. prising up to 180 projections can be recorded with a total dose as low as 5000 e nm−2. At higher magnifications with Exploiting the information contained in resolution targets in the 2 nm range, one aims not to ex- cryotomograms ceed a dose of 3000 e nm−2.The fraction of the dose that is Using automated procedures, cryotomograms can be spent on overhead [search, recentering, (auto)focusing] obtained more or less routinely within a few hours. Re- can be kept as low as 3% of the total; in other words, al- cording a tilt series and calculating a tomogram is indeed most all electrons are used for retrieving information.This less cumbersome and time consuming than going through has changed the perspectives of electron cryotomography the conventional procedures of plastic embedding and quite dramatically. It is now possible to study specimens sectioning the material. Once the tomogram is produced, embedded in vitreous ice as demonstrated with phantom it can be sectioned in silico in any desired direction and seg- cells (lipid vesicles encapsulating macromolecules) [12,13], mented for visualization (Fig. 2). With smaller structures organelles [14,15] or whole prokaryotic cells [16,17]. (e.g. bacteriophages docked onto liposomes), a resolution

http://www.trends.com S41 Review | A TRENDS Guide to Imaging Technologies Trends in Biotechnology Vol. 20 No. 8 (Suppl.), 2002

(a) (b)

Figure 2. An example of a of 2.5 nm has been obtained [13] and with whole Therefore, we prefer a different strategy, namely a tomographic reconstruction prokaryotic cells, resolution is in the range of 4.0–5.0 nm detection and identification based on structural signatures Electron tomographic reconstruction of the Archaeon according to different resolution criteria (Fourier trans- [21,22]. Provided that a high- or medium-resolution Methanospirillum hungatei form of small crystalline inclusions or Fourier ring corre- structure of the macromolecule of interest is available, this embedded in vitrified ice. The lation of single particle datasets extracted from tomograms). can be used as a template to perform a systematic search tomogram shown is reconstructed from 60 Further improvements in resolution can only be achieved of the reconstructed volumes for matching structures projections; the tilt increments with a finer 3D sampling; this, in turn, requires that the (Fig. 3). Such an approach is computationally demanding were 1.5° and the effective tilt range was ±45°. The cumulative specimen can be made more resistant to radiation damage because not only are the positions of the molecules un- dose for recording the whole (see later). known, but also are their orientations. Image simulation dataset was ~2000 e nm−2. Once a tomogram is obtained, it Beyond easily recognizable features, such as the multi- [13] and experimental studies using phantom cells can be ‘sectioned’ in silico in layered cell wall structure in Fig. 2, a cryotomogram of a (A.S. Frangakis et al., unpublished results) have shown that any direction desired. (a) Two orthogonal slices are displayed, cell contains an imposing amount of information. It is, in template matching is indeed a feasible approach and can one parallel to the long axis of fact, a 3D image of the entire proteome of the cell and it achieve a satisfactory level of fidelity. The search should the cell showing the depicts the whole network of macromolecular interactions be performed in an objective and reproducible manner multilayered cell wall and one perpendicular to it at the level of that provide the basis for higher cellular functions. However, and, therefore, should be entirely machine-based, not re- a ‘plug’; plugs are complex new strategies and innovative image analysis techniques quiring manual intervention. Ideally, the search is ex- paracrystalline protein arrays separating individual cells within are needed for ‘mining’ this information. Exploitation is haustive, detecting all copies of the target molecule, and it filaments. (b) The plug structure confronted with two major problems: cryotomograms are should be fast to enable the analysis of large datasets. and the flagella traversing it are also shown in an isosurface contaminated by substantial residual noise and distorted The perfection of such pattern recognition tools will be representation. The tomogram by missing data – in spite of optimized image acquisition crucial for taking full advantage of the power of electron has been ‘de-noised’ by nonlinear anisotropic diffusion and systems. ‘De-noising’ techniques, while improving the cryotomography [21]. manually contoured to highlight signal-to-noise ratio, modify the signal in a way that pre- the cell wall and the flagella [21] cludes quantitative post-processing. Moreover, the cyto- Ongoing instrumental developments and (the diameter of the displayed structure is ~450 nm). plasm is densely populated (‘crowded’) with molecules prospects literally touching each other [20]. It is therefore virtually With the current (non-isotropic) resolution of 4–5 nm, we impossible to perform a segmentation and feature extrac- can address only larger (>400 kDa) complexes in a cellular tion based on a visual inspection of the tomograms. In context. To widen the scope of cellular tomography we principle, one could find ways to introduce electron-dense need to overcome some of the present limitations. labels marking the positions of the molecules under scrutiny As illustrated in Fig. 4 for a lipid vesicle, distortions and facilitating their detection. Such an approach, however, and artifacts resulting from the ‘missing wedge’ can be re- would no longer be non-invasive and it would be difficult, duced quite significantly by dual-axis tilting combining if not impossible, to achieve quantitative detection. two single-axis tilt series with a discrete or continuous

S42 http://www.trends.com Trends in Biotechnology Vol. 20 No. 8 (Suppl.), 2002 A TRENDS Guide to Imaging Technologies | Review

Cross-correlation function (ϕ, θ, ψ, T) ϕ ψ

θ 1 ϕ 2 ψ 3

θ n

Volume Vin Templates (T) Parallel computation Volume Vout

TRENDS in Biotechnology

Figure 3. Strategy for detecting and identifying individual macromolecules in a tomogram Detection and identification of individual macromolecules in cellular tomograms is based on their structural signature.The crowded nature of the cytoplasm and ‘contamination’ with noise means that an interactive segmentation and feature extraction is not feasible. It requires sophisticated pattern recognition techniques to exploit the information contained in the tomograms. A volume-rendered presentation of a 3D image is presented on the left. Even though some high-density features might be visible, an unambiguous identification of individual structures would be difficult if not impossible given the residual noise. An approach that has been proven to work is based on template matching: templates of the macromolecules under scrutiny are obtained by a high- or medium-resolution technique (X-ray crystallography, NMR, electron crystallography or ). These templates (4× magnified in this figure) are then used to search the entire volume of the tomogram systematically for matching structures by cross-correlation, and the result is refined by multivariate statistical analysis. In principle, the 3D image has to be scanned for all possible Eulerian angles φ, ψ and θ around three different axes, with templates of all the different protein structures one is interested in [e.g. the thermosome (blue) and proteasome (yellow)]. The search procedure is computationally very demanding but can be parallelized with respect to the different angular combinations in a highly efficient way. Finally, the position and orientation of the different complexes can be mapped directly in the 3D image.

90° rotation around the z-axis [23,24]. We have recently chip of the camera. Currently, novel designs are being designed and built a dual-axis tilt cryoholder and tested it explored that exclude the adverse contribution of back- successfully (S. Nickell et al., unpublished results). scattered electrons, intrinsic to fiber coupling. Such lens- It is expected that keeping the specimen at (near) liquid coupled cameras might improve the detection efficiency helium temperature (~10K and below) instead of liquid ni- twofold. trogen temperature (~90–100K) will increase its radiation Taken together, these developments should help to resistance. Although measurements performed in the past push the limits of electron cryotomography and brighten came to quite divergent results regarding the protection the prospects to explore the uncharted territory of the factors, there is a broad consensus that the cryoprotection molecular architecture of the cytoplasm [28,29]. factor will be (at least) in the range of 2–3 [25]. While this Figure 4. Gain in information by dual-axis tilting might appear to be only a modest improvement, it would The schematic diagrams show help considerably in tomography; it would in fact enable us (a) (b) the sectors in the Fourier to double the number of projections, and the finer sampling domain, which remain unsampled owing to the limited in 3D should bring us closer to the 2 nm resolution target. tilt range (here ±60°). (a) In This, in turn, would enable the detection of smaller particles single-axis tilting there is a ‘missing wedge’, (b) in dual-axis in cellular volumes and improve the identification fidelity. tilting a ‘missing pyramid’. The With the new generation of TEMs, which can be operated artefacts resulting from the at both liquid nitrogen and liquid helium temperature, missing information are shown below for an ideal thin-walled a more accurate assessment of the protection factor will vesicle. (a) In the single-axis become possible. case, the vesicle walls perpendicular to the tilt axis and Improvements can be expected to come from optimizing the poles of the vesicle are detection systems for intermediate voltage (300–400 kV) poorly represented in the reconstruction. (b) In the microscopes [26,27]. With the slow-scan CCD cameras, two-axis case, only the poles of performance is affected by the discrete pixel size, electron the vesicle are affected by the missing information. In both (back) scattering, photon propagation and the (non- TRENDS in Biotechnology cases, the same threshold level optimal) coupling between the scintillator and the CCD has been applied.

http://www.trends.com S43 Review | A TRENDS Guide to Imaging Technologies Trends in Biotechnology Vol. 20 No. 8 (Suppl.), 2002

References 15 Mannella, C.A. (2001) Application of electron tomography to mitochondria research. Methods Cell Biol. 65, 245–256 1 Alberts, B. (1998) The cell as a collection of protein machines: 16 Grimm, R. et al. (1998) Electron tomography of ice-embedded preparing the next generation of molecular biologists. Cell 92, 291–294 prokaryotic cells. Biophys. J. 74, 1031–1042 2 Gavin, A-C. et al. (2002) Functional organization of the yeast proteome 17 Baumeister,W. et al. (1999) Electron tomography of molecules and by systematic analysis of protein complexes. Nature 415, 141–147 cells. Trends Cell Biol. 9, 81–85 3 Frank, J. (1992) Electron Tomography, Plenum Press 4 Crowther, R.A. et al. (1970) Three-dimensional reconstruction of 18 Dubochet, J. et al. (1988) Cryo-electron microscopy of vitrified spherical by Fourier synthesis from electron micrographs. specimens. Q.Rev.Biophys. 21, 129–228 Nature 226, 221–223 19 Grimm, R. (1996) Determination of the inelastic mean free-path in 5 Saxton,W.O. et al. (1984) The three-dimensional reconstruction of ice by examination of tilted vesicles and automated most probable imperfect two-dimensional crystals. Ultramicroscopy 13, 57–70 loss imaging. Ultramicroscopy 63, 169–179 6 Radon, J. (1917) Über die Bestimmung von Funktionen durch ihe 20 Ellis, R.J. (2001) Macromolecular crowding: obvious but Integralwerte längs gewisser Mannigfaltigkeiten. Berichte über die underappreciated. Trends Biochem. Sci. 26, 597–604 Verhandlungen der Königlich Sächsischen Gesellschaft der 21 Boehm, J. et al. (2000) Toward detecting and identifying Wissenschaften zu Leipzig. Math. Phys. Klasse 69, 262–277 macromolecules in a cellular context: template matching applied to 7 Hegerl, R. and Hoppe,W.(1976) Influence of electron noise on electron tomograms. Proc. Natl.Acad. Sci. U.S.A. 97, 14245–14250 three-dimensional image reconstruction. Z. Naturforsch. 31a, 22 Frangakis, A.S. and Hegerl, R. (2001) Noise reduction in electron 1717–1721 tomographic reconstructions using nonlinear anisotropic diffusion. 8 McEwen, B.F. et al. (1996) The relevance of dose-fractionation in J. Struct. Biol. 135, 239–250 tomography of radiation-sensitive specimens. Ultramicroscopy 60, 23 Mastronarde, D.A. (1997) Dual-axis tomography: An approach with 357–373 alignment methods that preserve resolution. J. Struct. Biol. 120, 9 Dierksen, K. et al. (1992) Towards automatic electron tomography. 343–352 Ultramicroscopy 40, 71–87 24 Penczek, P. et al. (1995) Double-tilt electron tomography. Ultramicroscopy 10 Dierksen, K. et al. (1995) Three-dimensional structure of lipid vesicles 60, 393–410 embedded in vitreous ice and investigated by automated electron 25 Chiu,W. et al. (1986) Cryoprotection in electron microscopy. J. Microsc. tomography. Biophys. J. 68, 1416–1422 141, 385–391 11 Koster,A.J. et al. (1997) Perspectives of molecular and cellular electron 26 Baumeister,W.and Herrmann, K-H. (1990) High-resolution electron tomography. J. Struct. Biol. 120, 276–308 microscopy in biology: sample preparation, image recording and 12 Grimm, R. et al. (1997) Energy-filtered electron tomography of processing. Biophys. Elect. Microsc. 5, 109–131 ice-embedded actin and vesicles. Biophys. J. 72, 482–489 27 Faruqi, A.R. and Subramaniam, S. (2000) CCD detectors in 13 Boehm, J. et al. (2000) FhuA-mediated phage genome transfer into high-resolution biological electron microscopy. Q.Rev.Biol. 33, 1–27 liposomes: a cryo-electron tomography study. Curr. Biol. 11, 28 McIntosh, J.R. (2001) Electron microscopy of cells: a beginning for a 1168–1175 new century. J. Cell. Biol. 153, F25–F32 14 Nicastro, D. et al. (2000) Cryo-electron tomography of neurospora 29 Schliwa, M. (2002) The evolving complexity of cytoplasmic structure. mitochondria. J. Struct. Biol. 129, 48–56 Nat. Rev.Mol. Cell Biol. 3, 291–296

S44 http://www.trends.com