REVIEWS TIBS 25 – DECEMBER 2000

624 0968 – 0004/00/$ – See front matter © 2000, Elsevier Science Ltd. All rights reserved. PII: S0968-0004(00)01720-5

OVER THE PAST 10 years, major strideshave been made towards harnessing theinstrumental resolving power of theelectron microscope for biological pur-poses1. At ,3 Å, this resolving powershould suffice to disclose the folds ofmacromolecules. One key step has beenthe recognition that preservation of na-tive structure is a performance-limitingfactor of paramount importance, andmeasures have been devised to preserveproteins through the ordeal of electronimaging. Destructive dehydration effectsare averted by immobilizing the speci-men in a thin layer of vitrified buffer byrapid freezing, then transferring it to acryogenic specimen holder for obser-vation – hence ‘cryo-electron micros-copy’. Damage from electron irradiationis limited by restricting exposures to theminimum needed for statistically de-fined images. The second pivotal inno-vation has been the pervasive use ofcomputers: to suppress noise by averag-ing over many individual images; to syn-thesize 3D density maps from 2D micro-graphs; to correct images for optical(phase-contrast) effects; and, to anincreasing extent, for data acquisition,

replacing human operation of the micro-scope by computer control.

To date, the fruits of this progress in-clude the derivation of atomic modelsfor several intrinsic membrane proteins,such as bacteriorhodopsin2 and light-harvesting complex3, at resolutions inthe 3.5–4.5-Å range, and visualization ofa dozen or so other membrane proteinsin three dimensions at resolutions of5–10 Å. An atomic model has also beenobtained for tubulin4, a major cytoskel-eton protein, and several other proteinsthat are not integral membrane com-ponents but nevertheless pack into well-ordered 2D lattices have also been stud-ied at somewhat lower resolution. Workon ‘single’ (i.e. noncrystalline) particleshas succeeded in depicting secondarystructure at resolutions in the range of7–10 Å, so far mainly for icosahedralviruses like the hepatitis B virus cap-sid5,6. In addition, many studies at res-olutions of 15–30 Å have yielded a wealthof information on domain and subdo-main organization, and intersubunitinteractions in various protein complexes.

As resolutions continue to improve,the question arises whether electron mi-croscopy will emulate or even supplantX-ray crystallography as a source ofhigh-resolution structures. Currently,this appears unlikely (but not imposs-ible); rather, the emerging relationshipis one of complementarity. Cryo-EM of-fers a promising approach for moleculesthat are refractory to X-ray diffractionanalysis, for example complexes thatare too large, unstable or insoluble, or

are available in insufficient amounts forcrystallographic studies to be feasible.Moreover, as outlined below, cryo-EMand X-ray crystallography – and to someextent, nuclear magnetic resonance(NMR) spectroscopy – can be combinedto overcome the limitations of eithermethod alone. Indeed, this hybrid ap-proach is emerging as a tool of choice forelucidating large complexes, includingthe macromolecular machines whosefunctional pre-eminence in a cellular con-text is becoming increasingly evident.

The aim of this article is to review themajor lines of development of 3D cryo-EM; to illustrate its potentialities withsome recent or otherwise notable exam-ples; and to assess its prospects for amajor role in the structural genomicsinitiative.

Electron crystallographyOrdered protein arrays, often only

one molecule thick, are too insubstan-tial to be analysed by X-ray diffractionbut make ideal specimens for a mode ofEM analysis that is crystallographic inspirit. The goal is to build up the 3DFourier transform of the repeating el-ement by recording data from arraystilted through various angles (Box 1).Each such view provides a central planethrough the 3D transform, obtained bycalculating the Fourier transform of thecorresponding digital image. As such, itprovides phases as well as amplitudes.For well-ordered crystals, electrondiffraction yields amplitudes that are su-perior to those from transformed images,and consequently are incorporated intothe transform. When all available datahave been integrated into a composite3D transform, the density map is ob-tained by inverse Fourier transformation.

This approach, electron crystallog-raphy, was pioneered on bacterio-rhodopsin2, a protein that forms or-dered arrays in the membrane of ahalophilic bacterium under conditionsof oxygen deprivation. The method isparticularly well suited for intrinsicmembrane proteins that are visualizedin their natural environment; that is, em-bedded in a lipid bilayer. While the num-ber of naturally crystalline membraneproteins is limited, a more widely appli-cable approach is to solubilize the mol-ecule of interest in detergent, purify it inmicelles, and reconstitute the proteinwith lipids, aiming to form 2D (or 3D)crystals7. This method of structuralanalysis is equally well applicable tomonolayer arrays of water-soluble pro-teins, such as those formed by tubulin

Macromolecular electronmicroscopy in the era of

structural genomics

Wolfgang Baumeister and Alasdair C. StevenMacromolecular machines carry out many cellular functions. Cryo-electronmicroscopy (cryo-EM) is emerging as a powerful method for studying thestructure, assembly and dynamics of such macromolecules, and theirinteractions with substrates. With resolutions still improving, ‘single-particle’ analyses are already depicting secondary structure. Moreover,cryo-EM can be combined in several ways with X-ray diffraction to enhancethe resolution of cryo-EM and the applicability of crystallography. Electrontomography holds promise for visualizing machines at work inside cells.

W. Baumeister is at the Dept of MolecularStructural Biology, Max-Planck-Institut fürBiochemie, Am Klopferspitz 18a, D-82152Martinsried, Germany; and A.C. Steven is atthe Laboratory of Structural Biology, NationalInstitute of Arthritis, Musculoskeletal andSkin Diseases, National Institutes of Health,Bethesda, MD 20892-2717, USA.Emails: baumeist@biochem.mpg.de;steven@calvin.niams.nih.gov

TIBS 25 – DECEMBER 2000

625

in the presence of Zn21 ions4, or proteinsarrayed on planar lipid monolayers8, aswith annexin9. Rational strategies havebeen devised to promote the associ-ation of proteins carrying high-affinitytags (e.g. oligo-His) with monolayers ofsuitably derivatized lipids, thus usingmembranes as crystallization templates10.

With large well-ordered arrays, elec-tron crystallography can yield atomicmodels, although the resolution tendsto be considerably lower in the dimen-sion perpendicular to the plane. Thisanisotropy occurs because it is notpracticable to tilt through a full 908, sothat part of the 3D transform – the ‘miss-ing cone’ – remains undefined (seeBox 1). Consequently, features organ-ized in the plane, such as the packing oftransmembrane a helices, tend to bewell represented, whereas features ar-ranged perpendicular to it, such asinterhelical linkers, are less well con-veyed. Advances in atomic force micro-copy now make it possible to delineate

REVIEWS

surface topography in submoleculardetail and compensate – at least in part– for this missing information11.Alternatively, with tubular membranescontaining helically ordered arrays ofproteins, as have been obtained withthe acetylcholine receptor12, and theCa21-ATPase13, the missing cone prob-lem is circumvented because (almost)all views of the repeating subunit arepresented in a single tube.

Electron crystallographic studies ofbacteriorhodopsin are now targetingdistinct states of the photocycle14. Anatomic model was obtained for a light-harvesting complex, revealing the spa-tial distribution of its prosthetic groups,in addition to the protein fold3. Recentwork on aquaporins15–17, an extensivefamily of water channels, has discov-ered a novel hairpin motif, with twocopies per subunit, penetrating themembrane from either side (Fig. 1). Alandmark result was the structure deter-mination of tubulin arrayed in planar

sheets4. Another protein of great inter-est that long resisted attempts to elicitstructural information below ~20-Å res-olution is connexin, the structural sub-unit of gap junctions. By exploiting well-ordered arrays obtained by expressingthe protein in HeLa cells, Unger et al.succeeded in visualizing the configur-ation of four transmembrane a helicesper subunit in the hexameric ring18.

‘Single-particle’ analysisFree-standing macromolecules offer

certain practical advantages for cryo-EM. Because there is no requirement forcrystallinity, virtually any particle is eli-gible providing that it exceeds a certainsize – currently at the 250–500-kDa level,depending on structure. For particlesthat assume random orientations in vit-rified films – which is usually the case –the ‘missing cone’ problem does notoccur, so that resolution should beisotropic (Box 1). Calculation of 3Dstructures from noncrystalline specimens

Box 1 Combining data to calculate 3D density maps in electron crystallography, single-particle analysis and electron tomography.

The 3D Fourier transform of the density distribution of the molecule of interest serves as a melting pot for combining cryo-electron microscopy(cryo-EM) data collected in each of three strategies: electron crystallography, single-particle analysis and tomography. These strategies are ap-propriate for different kinds of specimens. The key relationship exploited in these syntheses is that the Fourier transform of an EM image –which is a projection – gives a 2D transform that is a central plane of the full 3D transform, that is a plane that passes through its origin. Forease of illustration, the transforms are represented in Fig. I in two dimensions, so that this simplification relates to 1D projections and cen-tral lines in 2D transforms. In any case, the digital transforms are sampled on Cartesian lattices. The goal in each structural analysis is to fill in the values of the trans-

form at each point of this lattice as completely and accurately as possible.Then the density map can be obtained by simply calculating an inverseFourier transform. The resolution attained depends on the radius to whichreliable data extend in the transform (reciprocal space): the limits marked inFig. I represent the current state of the art for the three approaches.However, resolution is also reduced if parts of the transform remain empty;that is, if there are no data to be inserted. For instance, the ‘missing cone’(actually, in Fig. I, a missing wedge) is a consequence of being unable to tiltthrough a full 6908 and results in anisotropic resolution (see main text).Resolution is also limited by the amount of data missing in the spaces be-tween sampled planes; for instance, in tomography it is limited by the angu-lar increment between tilts (Dy, in radians), according to the relationship:r .(DDy)/2, where r is the resolution and D is the diameter of the object.Because micrographs are noisy, their Fourier transforms are also noisy andmust be averaged to suppress the noise (not possible in tomography, whereonly one image is obtained at each tilt angle). A rich variety of averagingstrategies have been developed to synthesize the available data. In electron crystallography, the transform is not sampled continuously, butonly at discrete points on each plane – the reciprocal lattice (shown in Fig. Ias bold points on each line). At these points, the signal-to-noise ratio ishighly favorable compared to corresponding points in the transforms of indi-vidual particle images, on account of implicit averaging. Each term in theFourier transform is a complex number. With large, well-ordered 2D crystals,electron diffraction provides superior amplitudes that can then be combinedwith phases taken from image transforms (an Argand diagram for one suchFourier component is shown in Fig. I, top right). In practice, many proceduresin current use do not literally fill up the 3D transform and calculate the in-verse transform: for instance, the data can be used to calculate the coeffi-cients of a functional expansion in Fourier space, or the density map can becalculated entirely by real space operations. However, these algorithms are,in effect, carrying out the same operations, and the scenario outlined hereallows a systematic comparison of the three approaches.

Ti BS

Electron crystallographyImages andelectron diffractionpatterns

Many tilted crystals

Single-particleanalysisImages of many(identical) singleparticles viewedin differentorientations

Untilted (or tilted)

Electron tomographyMultiple images of the same specimen recorded serially atdifferent tilt angles

CompositeFourier

transform

Missingcone

Missingcone

1/3.5

1/7

1/50

Densitymap

InverseFouriertransform

Amplitudefrom electron diffractogram;phase fromFT of image

Figure I

REVIEWS TIBS 25 – DECEMBER 2000

626

is widely called ‘single-particle’analysis. Commendably succinct,this term is arguably misleading be-cause the final 3D density map isderived not from a single particlebut from thousands of particles.However, it should be recalled thateach of these particle images hasto be classified on the basis of itsindividual information content.

The price to be paid for these ad-vantages is that the viewinggeometry of each particle must be determined before a 3D densitymap can be calculated. In general,the analysis proceeds in twostages. First, an initial density mapat moderate resolution, say 30 Å, issomehow acquired. A variety ofmethods can be followed to obtainit, including reconstruction of dataacquired by tilting the specimen, asin the ‘random conical tilt’method19 or electron tomography(see below); or without tilting, as in‘angular reconstitution’20; or byadopting a pre-existing 3D model ofa related complex, reconstructing atilt series of negatively staineddata21, or simply computer-gener-ating a model based on a priori in-formation. The second stage in-volves cyclic model-basedrefinement22. At each stage, moreaccurate values for the viewing an-gles of each particle are obtainedby matching it against a grid of re-projections of the current model,each of which represents a particu-lar view. Translational alignment isalso fine-tuned. Then, a refinedmap is calculated and this pro-cedure is iterated exhaustively.

The resolution attained depends onseveral factors, including the number ofparticles in the data set, the precision(and correctness) of the orientation andalignment parameters, and, of course,the quality of the original data. Evenunder the most favorable conditions,macromolecules yield noisy, low-con-trast images, and for particle orien-tations to be determined successfully,they must generate sufficient signal toallow reliable discrimination among thevarious projections. In practice, this re-quirement places a lower size limit onmacromolecules that can feasibly beanalysed – currently 250–500 kDa. Itmight be possible to extend this limitdownwards by binding heavy-metalclusters as fiducial markers, or by incor-porating the molecule of interest into alarger (thus more tractable) complex, or

to display it on some larger entity suchas a viral capsid23,24. Complexes of1–10 MDa are optimal for high-resolutionsingle-particle analysis. This approachdoes not appear to be subject to anupper mass limit, at least for complexeswhose structures are deterministic tothe atomic level, although depth of focuseventually becomes a limiting factor.

The first single-particle analyses atresolutions below 10 Å were of the cap-sids of hepatitis B virus (HBV)5,6 (Fig. 2)and papillomavirus25. Icosahedral cap-sids have the advantage of 60-fold sym-metry, which reduces the number ofparticles required to obtain a desiredamount of averaging, but otherwise theyoffer the same advantages (and tribu-lations) as other single particles. As 10 Åis the spacing typical of close-packed a helices, density maps in the sub-10-Å

range should be particularly infor-mative for proteins with high a-helical contents, because theyshould reveal the relative disposi-tion and, in favorable cases, theconnectivity of the helices. HBVcapsid protein is such a molecule,and the main part of the fold pro-posed on the basis of cryo-EM(Ref. 6) – a novel one for a capsidprotein – was confirmed and theremainder corrected by cryo-EMlocalization of fiducial markers(summarized in Ref. 26) and X-raycrystallography27. These studieshave been followed by maps of thehuge herpes simplex virus cap-sid28, revealing some 20 potentiala-helical segments in its 150-kDamajor capsid protein, and ofSemliki Forest virus29, an en-veloped virus, revealing the trans-membrane portions and spikelikeprojections of its glycoproteins.Recently, a comparably detaileddensity map was reported for aribosomal subunit30, and prospectsare excellent for rapid growth inthe number of analyses in thisresolution regime and beyond.

Electron tomographyThe most general method for ob-

taining 3D information by EM is to-mography31; this is applicable notonly to isolated particles but alsoto pleomorphous structures likemitochondria and other organ-elles, whole cells, or tissue sec-tions. The basic concept is notnew, but formidable technical ob-stacles prevented realization of itspotential until recently. The key

problem lay in reconciling two conflict-ing requirements: to obtain a detailedand undistorted reconstruction, a tilt se-ries must be recorded that covers aswide an angular range as possible in asmany increments as possible (Box 1); atthe same time, the electron dose mustbe kept subcritical if radiation damageis not to erase important details. In prin-ciple, one could partition a fixed totaldose over as many projections as are re-quired; however, the resulting projec-tions would be correspondingly noisier,making alignment of them for recon-struction unreliable. This problem isaggravated by the limited mechanicalaccuracy of specimen holders. With theadvent of computer-controlled micro-scopes and large-area digital cameras, it became possible to implement auto-mated procedures for the acquisition of

Figure 1(a) Three-dimensional density map of human aquaporin1 (AQP1) water channel, determined at 4 Å resolution byelectron crystallography of ice-embedded 2D crystals17.The view is normal to the bilayer and the chicken-wireand solid densities are at 1.5 and 2 sigma, respec-tively. Scale bar, 10 Å. (b) Stereo view of the threadingof AQP1 polypeptide chain across the lipid bilayer mem-brane. The six tilted transmembrane a helices (in solidcolor) in a monomer form a barrel enclosing the water-selective pathway. This pathway is also lined by two in-terhelix loops (in solid white) that contain short a heli-ces and house a pair of functionally important, highlyconserved NPA tripeptide motifs. Courtesy of A. Mitra.Reproduced, with permission, from Ref. 17.

TIBS 25 – DECEMBER 2000

627

REVIEWSfact, with less experimentaleffort. In general, however,the latter approach needs anunbiased and sufficiently de-tailed starting model, and thisis where the two approachescan be combined most effec-tively. Electron tomographycan provide reliable startingmodels, which can then be re-fined and extended by single-particle analysis33.

Hybrid crystallography: themarriage of X-ray diffraction andcryo-EM

Visualizing interactions.Cryo-EM, while not yet yield-ing individual structures atresolutions as high as thoseobtained with X-ray crystal-lography, is readily applica-ble to macromolecular com-plexes. If the structures ofthe individual componentsare known to high resolution,a cryo-EM density map of theentire complex, in the inter-mediate resolution range of15–30 Å, will usually allowtheir relative positions andorientations to be specifiedto within a few angstroms,thus identifyingthe interactionsurfaces. This cru-cial informationabout interactionscannot, in general,be inferred fromthe structures ofthe individualcomponents, evenat very high reso-

lution, and crystals of func-tional complexes tend to bea rather rare commodity.

The first example of suchwork described the dockingof a monoclonal Fab onto rhi-novirus, pinpointing its epi-tope on the capsid surface34,subsequently confirmed bycrystallography35. Shortlythereafter, the myosin S1ATPase was fitted into acryo-EM map of S1-deco-rated actin filaments, depict-ing a rigor cross-bridge36. Athird success came from de-termining the placement ofadenovirus hexons – trimersof 109-kDa subunits – on thevirion surface37. Variations

on this theme have since proliferated instudies of the cytoskeleton38, amongother topics.

In most such analyses, the dockinghas been done by hand, using moleculargraphics. The EM density map is repre-sented as a transparent surface, and astick model of the molecule to bedocked is manipulated to optimize thefit; that is, to minimize excursions of themolecule beyond the surface and unoc-cupied lacunae inside it. However, theneed for quantitative, objective fittingprocedures has been recognized39,40. Inparticular, it will be valuable to imposecalibrated figures of merit for the fits,and to develop criteria for whether dis-crepancies reflect shortcomings of thecryo-EM map or the subunit model (e.g.in the crystal structure vis-à-vis thesolution structure41), or are indicative of conformational changes induced informing the complex. To confirm the ori-entations of components in complexes,heavy-metal clusters can serve as valu-able site-specific markers42.

Phasing X-ray crystallographic datasets with EM-derived molecular en-velopes. Diffraction-quality crystals oflarge complexes call for different phas-ing methods than are applicable to crys-tals of typically sized globular proteins.For instance, heavy atoms might lack

tomographic data sets. Rather complexprotocols now allow correction for theimperfections of the experimental set-up, while minimizing the cumulativeelectron dose. Data sets of 100–200 pro-jections can be recorded at doses lowenough not to visibly damage ice-embedded specimens.

As applied to single particles, themethod has produced a 3D structure of the archaeal thermosome, a hexa-decameric chaperone, that was goodenough to allow docking of its domainsof known structure32 (Fig. 3). Tomo-graphic density maps of numerous indi-vidual particles were calculated,screened to remove outliers, and aver-aged. In this particular case the sameresult could, no doubt, have been ob-tained by single-particle methods – in

Figure 2(a) The capsid of hepatitis B virus as determined by3D reconstruction to 9 Å resolution5; and (b) a molec-ular model26 of the dimer of 149-residue subunits thatserves as its building block. The sites of the three ref-erence points marked – the N terminus, the C termi-nus and an internal peptide (residues 78–83) – weredetermined by high-resolution cryo-electron microscopy(cryo-EM) labeling techniques, as summarized inRef. 26. The C-terminal peptide, residues 140–149, isan important determinant of capsid size, namely inspecifying the relative amounts of 180-subunit and240-subunit capsids that are assembled. It is also theattachment point of the basic domain, residues150–183, that binds nucleic acid inside capsids madeof the full-length protein. Courtesy of J. Conway.

Figure 3Three-dimensional map of the archaeal thermosomeholoenzyme as reconstructed from cryo-electron mi-crographs. The thermosome is a hexadecamericchaperone, and the crystal structures of the subunitsare modelled into the molecular envelope32.Reproduced, with permission, from Ref. 32.

REVIEWS TIBS 25 – DECEMBER 2000

628

the scattering power to alter the struc-ture factors sufficiently to allow solutionby isomorphous replacement. Alterna-tively, such crystals might yield to aphase extension approach based on ex-ploiting molecular envelopes defined bycryo-EM, supplemented by noncrystal-lographic symmetry. If the number ofcomplexes and their positions in theunit cell are known, the envelope speci-fies those portions of the unit cell whosedensity is at a constant level. This pro-vides a ‘solvent flattening’ constraint,which – with the additional constraint ofnoncrystallographic symmetry – is im-posed iteratively, starting with a low-resolution model (e.g. the molecular en-velope from cryo-EM). This model isprogressively embellished by introduc-ing diffraction data to higher and higherresolution, and reimposing the con-straints at each cycle.

To date, this approach has workedwell with icosahedral viruses with 5–3–2symmetry, for example Norwalk virus43

and bacteriophage HK9744, the latterstudy revealing a novel capsid proteinfold; and for ClpP, a barrel-like proteaseconsisting of two heptameric rings with7–2 symmetry45. The method appears tohave much potential for the analysis oflarge complexes. A line of enquiry bear-ing on its applicability is to determinethe minimum constraints needed forsuch phasing schemes to converge tostable, correct solutions, and the extent

to which the requirement for a highorder of symmetry might be relaxed orcompensated by improved resolution incryo-EM density maps. Work on theribosome – which lacks internal symme-try – has demonstrated that EM-derivedinformation can be usefully exploited in crystallographic studies of suchparticles46.

Visualizing ‘excited’ states ofmacromolecular complexes. The reac-tion cycles and assembly pathways ofmacromolecular complexes often in-volve passaging through short-lived,structurally distinct states, concomitantwith the binding and processing of lig-ands or cofactors. Elucidation of suchelaborate phenomena requires struc-tural knowledge of each step in the path-way. If a complex is amenable to crystal-lization, it is likely to be captured in itsmost stable ‘ground state’, whereas theprecursor or transitional states, as ‘ex-cited states’, can be more elusive47. Itcan also happen that salting out of pro-tein complexes into crystals fixes themin some nonphysiological configura-tion48. However, intermediate states ofcomplexes in solution can be arrestedby cryo-EM, using rapid-freezing49 or byadjusting buffer conditions to extendthe lifetime of a given state or its preva-lence in mixed populations50. In such sit-uations, an atomic model of the groundstate may be adapted into pseudo-atomic models of transitional states by

assuming that the main structural el-ements – domains or subunits – are con-served, but undergo rigid-body move-ments. This assumption reduces thenumber of free parameters describingthe intermediate state to the point thatthey can be stably determined from acryo-EM reconstruction at moderateresolution. However, the possibility re-mains that such transitions might alsoinvolve local refolding or remodeling,and this eventuality can ultimately beaddressed only by improving the resolu-tion of the reconstruction or obtaining acrystal structure.

Among other applications, this ap-proach – integrating cryo-EM and X-raycrystallography – has been used to in-vestigate the ‘cell-entry intermediate’states of poliovirus (Fig. 4). Comparedwith the native virus, these particles ex-hibit sizeable movements of their b-sandwich domains that create gaps be-tween the capsid proteins throughwhich RNA and internal peptides mightexit51. Another incisive example hasbeen the domain movements elicitedwhen the chaperone GroEL switches be-tween nucleotide states52. This com-bined approach has great potential in,among other topics, structural studiesof ribosomes in which atomic modelsare emerging from crystallography (e.g.Ref. 53). Cryo-EM is well placed to ex-ploit this information in elucidating thetranslation cycle by visualizing ribo-somes interacting with elongation andinitiation factors, nascent peptides, etc.(e.g. Refs 54,55) (Fig. 5).

Cryo-EM and structural genomicsThe goal of structural genomics is to

determine the structures of the proteinsidentified in genome sequences, ordomains thereof; in particular, to solverepresentatives of each topologicallydistinct fold. With so many targets, em-phasis has been put on automation toachieve high throughput in expressionand crystallization trials. Doubtlessly,structural genomics will yield many do-main structures, both because domainstend to be easier to crystallize than in-tact proteins and because they are of asize amenable to NMR. Although thismassive repository of structural infor-mation will surely contribute to solvingthe protein folding problem, its implica-tions for a mechanistic understanding ofbiological functions are harder to pre-dict, as most cellular functions arecarried out, not by individual domainsbut by large complexes56. The latter en-tities do not readily lend themselves to

Figure 4Binding of poliovirus to its receptor (Pvr)23,24. Left: cryo-electron microscopic (cryo-EM) recon-struction of poliovirus (red) with the water-soluble ectodomain of Pvr (green) stoichiometricallybound23. In addition to revealing the binding interaction, the reconstruction aided the construc-tion of a quasi-atomic model of Pvr (right), previously known to consist of three immunoglobu-lin-like domains. Domains 1 (which contacts the virus) and 2 are approximately aligned,whereas domain 3, which is closest to the cell surface, is offset at an angle of ~60°. The re-construction clearly shows two carbohydrate moieties (arrows) attached to arginine residueson domain 2. Courtesy of D. Belnap.

TIBS 25 – DECEMBER 2000

629

high-throughput tactics because meth-ods do not yet exist for mass productionby co-expression and assembly of largenumbers of gene products. Thus there isa distinct need for methods that yieldoverall structures of large complexesthat might be available only in limitedsupply, or detailed structures of otherclasses of automation-unfriendly macro-molecules such as membrane proteins.

In this context, further technicalprogress coupled with systematic inte-gration with bioinformatics should allowcryo-EM to emerge as a major player instructural genomics. There are two as-pects to be considered: first, acceleratedexploitation of state-of-the-art methods;and second, further extension of resolu-tion. In either event, desiderata include(for single-particle analysis): automationof data collection57; streamlining andstandardization of analysis procedures;and acquisition of primary data from themost powerful microscopes currently inoperation. To make these resources avail-able, there appears to be a persuasivecase for establishing a number of cutting-

REVIEWS

edge installations to serve the cryo-EMcommunity in much the same way as syn-chrotron beamlines meet the needs ofcrystallographers.

Computational advances will also beessential, and further refinement, inte-gration and standardization of data anal-ysis must be achieved in order to distillthe information from the very large datasets that are needed, for example>105–106 particles. Here, parallel com-puting58, which lends itself to this appli-cation, should be a major factor. A goalof ~4.5-Å resolution in single-particleanalyses appears realistic, for at leastsome classes of molecules, and shouldsuffice to resolve strands in b sheetsand the connectivity of secondary struc-ture elements, and consequently, thebackbone folds of these proteins.

Challenges for bioinformatics: en-hancing resolution. Exciting as theabove prospects are, it appears unlikelythat cryo-EM of single particles will ad-vance in the short term to resolutions atwhich side-chain configurations and thebinding of water molecules can be as-

signed (say, 2–2.5 Å). In this context, anintriguing proposition is whether itmight be possible to draw empiricallyon the growing database of high-resol-ution X-ray structures to refine a struc-ture solved at, say, 4.5-Å resolution inwhich the path followed by the backboneis mapped out, to one at 2.5 Å in whichthe stereochemistry is more fully de-fined. Although certainly challenging,this might turn out to be a more achiev-able goal than de novo structural predic-tion from amino acid sequences.

A second attractive possibility is todevelop methods that scan the databaseof high-resolution domain structures todiscriminate domains in EM-deriveddensity maps at the currently achievableresolutions of 7–15 Å, thereby allowingthe maps to be interpreted in muchgreater detail than these numbers wouldimply. In this notion, the problem ofstructural determination is reduced toone of pattern recognition in which theelements to be recognized are domains,not amino acids. Because domains aremuch larger than amino acids, and the

Stalk

Head

Spur

L1

Platform

EF-G

Channel

EF-G

BodyArc

L1

Spur

Stalk

Central protuberance

Head

30S

50S

Channel

Body

Stalk base

Platform origin

(a) (b)

Figure 5Interactions of elongation factors with the 70S ribosome of Escherichia coli. (a) Binding of EF-G in the presence of fusidic acid55. Left column: 20-Åmap of the ribosome (blue) and the observed difference map due to EF-G (red) in two views. The top view shows the protrusion formed by domain IVof EF-G reaching into the decoding site, thereby displacing the A-site tRNA. The bottom view shows an arc-like contact between EF-G and the ribosome,involving the G9 domain of EF-G. Right column: corresponding transparent views of the ribosome with a model of the six color-coded domains of EF-G,obtained by fitting the X-ray structure into the difference map. Domain I(G) – magenta; I(G9) – brown; II – blue; III – green; IV – yellow; V – red. The fit-ting required a rotation of domains III, IV and V, indicating a conformational change in EF-G upon binding to the ribosome. Courtesy of J. Frank.Reproduced, with permission, from Ref. 55. (b) The ternary complex (EF-Tu/tRNA/GTP) on the kirromycin-stalled 70S E. coli ribosome at 13 Å reso-lution (H. Stark, M. Rodnina, F. Zemlin, W. Wintermeyer and M. van Heel, unpublished): 30S subunit, gold (left); 50S subunit, brown (right). The ternarycomplex-associated density consists of: domain 1 of EF-Tu – red; domain 2 of EF-Tu – pink; domain 3 of EF-Tu – yellow; tRNA – blue; P-site-bound tRNA– green. The ternary complex and the P-site-bound tRNA are semitransparent (relatively high level of transparency). The PDB coordinates of the ternarycomplex have been modified. Courtesy of H. Stark.

REVIEWS TIBS 25 – DECEMBER 2000

630

number of distinct domain folds is notinordinate, they should, in principle, berecognizable in density maps at muchlower resolution.

Visualizing molecules inside cells. Ithas long been a dream of biologists tosee the molecular architecture of livingcells. In addition to robust machines likethe ribosome, there are undoubtedlymany ‘teams’ of interacting moleculesperforming cellular functions, that areheld together by forces too weak to sur-vive biochemical isolation56. To this end,recent developments in automated elec-tron tomography have made it possibleto obtain 3D reconstructions of wholeice-embedded prokaryotic cells59 ororganelles such as mitochondria60,61,with resolutions of 60–80 Å. Furtherprogress with instrumentation shouldenter the realm of molecular resolution(20–40 Å), thus bridging the gap betweenmolecular and cellular structural biology.

For reasons noted above, the objec-tive of cellular electron tomographycannot be to image individual macro-molecules at high resolution. Given thecrowding of most compartments andthe inevitably low signal-to-noise ratioof tomograms, it will generally not bepossible to interpret them in terms ofsupramolecular architectures by visualinspection. Nevertheless, if high-to-medium-resolution structures of themolecules under scrutiny are available,they could be used as templates to scanthe reconstructed volume and thus mapout the distribution of matching struc-tures. Here, the goal is to identify macro-molecules by virtue of their structuralsignatures and thereby to locate them ina cellular environment. The searchshould, of course, be performed withmultiple templates, revealing not onlyterritorial distributions, but also spatialrelationships in functional ‘neighbor-hoods’. Once the position and orien-tation of a complex have been deter-mined, subvolumes can be extracted,classified and averaged, providinginsight into intracellular interactions. Arecent study has demonstrated thatsuch an approach is feasible, albeit com-putationally demanding62. The methodappears to be sufficiently sensitive todiscriminate, with good fidelity, mol-ecular assemblies of similar size andshape in simulated and real tomographicdata sets.

ConclusionsThere is growing awareness that

many key cellular functions are performedby ensembles of macromolecules –

molecular machines. Whereas some ma-chines are abundant and robust, like the‘monolithic’ ribosome53, others are pre-sent in only a few copies per cell or aretoo delicate to survive isolation proce-dures, ruling out X-ray crystallographyor NMR spectroscopy as means to deter-mine their structure. Also, during theirreaction cycles, machines pass throughshort-lived altered states that will be, atbest, difficult to capture by crystallog-raphy. In this context, there is immenseopportunity for cryo-EM, especially asboosted by merging crystallographicstructures of individual subunits intomoderate-resolution cryo-EM densitymaps of whole complexes. Electrontomography has now advanced to thepoint where it is a realistic goal toglimpse molecular machines operatinginside cells: again, value will be added by integrating higher-resolution dataobtained on isolated complexes.

AcknowledgementsWe thank J. Bohm for help with graph-

ics, and D. Rees and P. Bjorkman forhelpful comments.

References1 Glaeser, R.M. (1999) Electron crystallography: present

excitement, a nod to the past, anticipating the future.J. Struct. Biol. 128, 3–14

2 Henderson, R. and Unwin, P. N. (1975) Three-dimensional model of purple membrane obtained byelectron microscopy. Nature 257, 28–32

3 Kuhlbrandt, W. et al. (1994) Atomic model of plantlight-harvesting complex by electron crystallography.Nature 367, 614–621

4 Nogales, E. et al. (1998) Structure of the alpha betatubulin dimer by electron crystallography. Nature 391,199–203

5 Conway, J.F. et al. (1997) Visualization of a 4-helixbundle in the hepatitis B virus capsid by cryo-electronmicroscopy. Nature 386, 91–94

6 Boettcher, B. et al. (1997) Determination of the fold ofthe core protein of hepatitis B virus by electroncryomicroscopy. Nature 386, 88–91

7 Hasler, L. et al. (1998) 2D crystallization of membraneproteins: rationales and examples. J. Struct. Biol.121, 162–171

8 Uzgiris, E.E. and Kornberg, R.D. (1983) Two-dimensional crystallization technique for imagingmacromolecules, with application to antigen-antibody-complement complexes. Nature 301, 125–129

9 Olofsson, A. et al. (1994) Two-dimensional structure ofmembrane-bound annexin V at 8 Å resolution. J. Struct.Biol. 113, 199–205

10 Kubalek, E. et al. (1994) Two-dimensionalcrystallization of histidine-tagged HIV-1 reversetranscriptase promoted by a novel nickel-chelating lipid.J. Struct. Biol. 113, 117–123

11 Heymann, J.B. et al. (1999) Charting the surfaces ofthe purple membrane. J. Struct. Biol. 128, 243–249

12 Miyazawa, A. et al. (1999) Nicotinic acetylcholinereceptor at 4.6 Å resolution: transverse tunnels in thechannel wall. J. Mol. Biol. 288, 765–786

13 Zhang, P. et al. (1998) Structure of the calcium pumpfrom sarcoplasmic reticulum at 8-Å resolution. Nature392, 835–839

14 Subramaniam, S. and Henderson, R. (2000) Molecularmechanism of vectorial proton translocation bybacteriorhodopsin. Nature 406, 653–657

15 Li, H. et al. (1997) Molecular design of aquaporin-1water channel as revealed by electron crystallography.Nat. Struct. Biol. 4, 263–265

16 Murata, K. et al. (2000) Structural determinants ofwater permeation through aquaporin-1. Nature 407,599–605

17 Ren, G. et al. (2000) Three-dimensional fold of thehuman AQP1 water channel determined at 4 Åresolution by electron crystallography of two-dimensional crystals embedded in ice. J. Mol. Biol.301, 369–387

18 Unger, V.M. et al. (1999) Three-dimensional structure ofa recombinant gap junction membrane channel.Science 283, 1176–1180

19 Radermacher, M. (1988) Three-dimensional reconstructionof single particles from random and nonrandom tilt series.J. Electron Microsc. Tech. 9, 359–394

20 van Heel, M. (1987) Angular reconstitution: a posterioriassignment of projection directions for 3Dreconstruction. Ultramicroscopy 21, 111–124

21 Kolodziej, S.J. et al. (1997) Utility of Butvar supportfilm and methylamine tungstate stain in three-dimensional electron microscopy: agreement betweenstain and frozen-hydrated reconstructions. J. Struct.Biol. 120, 158–167

22 Baker, T.S. and Cheng, R.H. (1996) A model-basedapproach for determining orientations of biologicalmacromolecules imaged by cryoelectron microscopy. J. Struct. Biol. 116, 120–130

23 Belnap, D.M. et al. (2000) Three-dimensional structureof poliovirus receptor bound to poliovirus. Proc. Natl.Acad. Sci. U. S. A. 97, 73–78

24 He, Y. et al. (2000) Interaction of the poliovirus receptorwith poliovirus. Proc. Natl. Acad. Sci. U. S. A. 97, 79–84

25 Trus, B.L. et al. (1997) Novel structural features ofbovine papillomavirus capsid revealed by a threedimensional reconstruction to 9Å resolution. Nat.Struct. Biol. 4, 13–20

26 Conway, J.F. et al. (1998) Localization of the N terminusof hepatitis B virus capsid protein by peptide-baseddifference mapping from cryoelectron microscopy. Proc.Natl. Acad. Sci. U. S. A. 95, 14622–14627

27 Wynne, S.A. et al. (1999) The crystal structure of thehuman hepatitis B virus capsid. Mol. Cell 3, 771–780

28 Zhou, Z.H. et al. (2000) Seeing the herpesvirus capsidat 8.5 Å. Science 288, 877–880

29 Mancini, E.J. et al. (2000) Cryo-electron microscopyreveals the functional organization of an envelopedvirus, Semliki Forest virus. Mol. Cell 5, 255–266

30 Matadeen, R. et al. (1999) The Escherichia coli largeribosomal subunit at 7.5 Å resolution. Structure Fold.Des. 7, 1575–1583

31 Koster, A. and Agard, D. (eds) (1997) Special issue onelectron tomography. J. Struct. Biol. 120(3).

32 Nitsch, M. et al. (1998) Group II chaperonin in an openconformation examined by electron tomography. Nat.Struct. Biol. 5, 855–857

33 Walz, J. et al. (1999) Capsids of tricorn protease studiedby electron cryomicroscopy. J. Struct. Biol. 128, 65–68

34 Wang, G.J. et al. (1992) Identification of a Fabinteraction footprint site on an icosahedral virus bycryoelectron microscopy and X-ray crystallography.Nature 355, 275–278

35 Smith, T.J. et al. (1996) Neutralizing antibody to humanrhinovirus 14 penetrates the receptor-binding canyon.Nature 383, 350–354

36 Rayment, I. et al. (1993) Structure of the actin-myosincomplex and its implications for muscle contraction.Science 261, 58–65

37 Stewart, P. L. et al. (1993) Difference imaging ofadenovirus: bridging the resolution gap between X-raycrystallography and electron microscopy. EMBO J.12, 2589–2599

38 Steinmetz, M.O. et al. (1998) An atomic model ofcrystalline actin tubes: combining electron microscopywith X-ray crystallography. J. Mol. Biol. 278, 703–711

39 Volkmann, N. and Hanein, D. (1999) Quantitative fittingof atomic models into observed densities derived byelectron microscopy. J. Struct. Biol. 125, 176–184

40 Wriggers, W. et al. (1999) Situs: A package for dockingcrystal structures into low-resolution maps fromelectron microscopy. J. Struct. Biol. 125, 185–195

41 Yu, X. and Egelman, E.H. (1992) Structural data suggestthat the active and inactive forms of the RecA filament arenot simply interconvertible. J. Mol. Biol. 227, 334–346

42 Kikkawa, M. et al. (2000) 15 Å resolution model of themonomeric kinesin motor, KIF1A. Cell 100, 241–252

43 Prasad, B.V. et al. (1999) X-ray crystallographic structureof the Norwalk virus capsid. Science 286, 287–290

44 Wikoff, W.R. et al. (2000) Topologically linked rings ofcovalently joined protein subunits form the dsDNAbacteriophage HK97 capsid. Science 289, 2129–2133

45 Wang, J. et al. (1997) The structure of ClpP at 2.3 Åresolution suggests a model for ATP-dependentproteolysis. Cell 91, 447–456

46 Ban, N. et al. (1998) A 9 Å resolution X-raycrystallographic map of the large ribosomal subunit.Cell 93, 1105–1115

TIBS 25 – DECEMBER 2000

6310968 – 0004/00/$ – See front matter © 2000, Elsevier Science Ltd. All rights reserved. PII: S0968-0004(00)01714-X

THE CELL CAN be conceived as a bio-chemical information-processing devicewhose response to the environment de-pends on the state of spatially organizednetworks of protein activities. Localizedprotein interactions, covalent modifi-cations, proteolytic processing and con-formational changes are the weighted in-formation transfer functions by whichthe state of these complex molecular net-works is determined. The interconnectiv-ity and spatial organization of proteinsystems that sustain basic cellular func-tions is only maintained in the context ofthe whole intact molecular architectureof the cell. It is clear that understandingcell function by integrating molecular ac-tivities with spatial organization is an im-portant challenge for modern biology.

Numerous biochemical assays revealingprotein modifications, interactions ortransport have already helped to bringus closer to this goal. However, none ofthe methods that are based on reconsti-tuted systems in the test-tube can fullytake into account the compartmentalizedand interconnected nature of these reac-tions in cells. It is therefore imperative todevelop methods that can measure thedynamics of these biochemical reactionsin the intact cell and thereby extract thespatial organization in vivo. Only re-cently, and most probably because of theavailability of genetically encoded fluo-rescent proteins, has it become possibleto image not only cellular processes suchas protein translocation or transport butalso basic reactions such as proteininteractions, proteolysis and phos-phorylation in intact cells.

Fluorescent proteins for live cell imagingFluorescence imaging of biological

reactions in living cells requires tech-

niques to fluorescently label the macro-molecules involved. These fluorescentprobes need to be specific, interfere aslittle as possible with the reactions to bevisualized, must not perturb the physio-logical conditions of the cells, and musthave favourable spectroscopic proper-ties for efficient detection. For manyyears the only way to study a fluor-escent protein in living cells was to purify,chemically modify and then (re)intro-duce it into living cells by, for example,capillary microinjection1. Unfortunately,the widespread use of this approach hasbeen inhibited by technical limitations.For example, it is often difficult to labellarger proteins fluorescently becausethey are more difficult to express andpurify in sufficient amounts by recombi-nant methods. Furthermore, trans-membrane or luminal proteins of thesecretory pathway are difficult or im-possible to label because the injected,fluorescent proteins do not translocateinto the lumen of the endoplasmicreticulum or the endogenous protein isnot accessible to fluorescent antibodies.Therefore, one of the most exciting ad-vances in fluorescent in vivo labellingtechniques has been the discovery ofthe green-fluorescent protein (GFP) andits spectral variants2–4. Because GFP isgenetically encoded and emits visiblefluorescence without the need for cofac-tors, it has been used extensively as afluorescent tag fused to cloned proteins.In principle, any cDNA of interest can befused to the coding sequence of GFP andexpressed as a fluorescent protein incells. In this way even membrane or lu-minal proteins can be labelled in vivo.For these reasons, the number of appli-cations of GFP in combination withfluorescence time-lapse microscopy hasbeen enormous. This has helped to-wards a better understanding of the dy-namics of the organization and architec-ture of living cells or organisms3,5,6.

Although GFP tagging of proteins hasclear advantages, the GFP technology

REVIEWS

Observing proteins in theirnatural habitat: the living cell

Philippe I.H. Bastiaens and Rainer PepperkokFluorescence microscopy has played a tremendous role in uncovering themorphological features of cells and the expression pattern of proteins byimmunofluorescence. Since the discovery of green-fluorescent proteins(GFPs), this technique has undergone a revival in the life sciences as thespatial distribution of ectopically expressed fusion proteins inside livingcells can now be followed more easily. By further exploiting the photophysi-cal properties of the emitted fluorescence with microspectroscopic meth-ods, spatial information on the biochemical parameters of intracellular pro-cesses and reactions can be obtained. This possibility will not only play animportant role in the understanding of biochemical reactions in signal pro-cessing and fidelity but also help to uncover the molecular mechanisms oforganelle and cell morphogenesis.

P.I.H. Bastiaens and R. Pepperkok are atEMBL Heidelberg, Meyerhofstr. 1,69117 Heidelberg, Germany. Emails: bastiaen@embl-heidelberg.de;pepperko@embl-heidelberg.de

47 Dokland, T. et al. (1997) Structure of a viralprocapsid with molecular scaffolding. Nature389, 308–313

48 Gutsche, I. et al. (2000) Conformationalrearrangements of an archaeal chaperonin uponATPase cycling. Curr. Biol. 10, 405–408

49 Walker, M. et al. (1995) Millisecond time resolutionelectron cryo-microscopy of the M-ATP transient kineticstate of the acto-myosin ATPase. Biophys. J. 68,87S–91S

50 Lata, R. et al. (2000) Maturation dynamics of a viralcapsid: visualization of transitional intermediatestates. Cell 100, 253–263

51 Belnap, D.M. et al. (2000) Molecular tectonic modelof virus structural transitions: the putative cell entrystates of poliovirus. J. Virol. 74, 1342–1354

52 Roseman, A.M. et al. (1996) The chaperonin ATPasecycle: mechanism of allosteric switching andmovements of substrate-binding domains in GroEL.Cell 87, 241–252

53 Ban, N. et al. (2000) The complete atomic structure ofthe large ribosomal subunit at 2.4 Å resolution.Science 289, 905–920

54 Stark, H. et al. (1997) Visualization of elongationfactor Tu on the Escherichia coli ribosome. Nature389, 403–406

55 Agrawal, R.K. et al. (1998) Visualization of elongationfactor G on the Escherichia coli 70S ribosome: themechanism of translocation. Proc. Natl. Acad. Sci. U. S. A. 95, 6134–6138

56 Special issue on macromolecular machines. (1998)Cell 92, 291–390

57 Carragher, B. et al. Leginon: an automated system foracquisition of images of vitreous ice specimen. J. Struct. Biol. (in press)

58 Martino, R.L. et al. (1994) Parallel computing inbiomedical research. Science 265, 902–908

59 Baumeister, W. et al. (1999) Electron tomography ofmolecules and cells. Trends Cell Biol. 9, 81–85

60 Frey, T.G. and Mannella, C.A. (2000) The internalstructure of mitochondria. Trends Biochem. Sci. 25,319–324

61 Nicastro, D. et al. (2000) Cryo-electron tomography ofNeurospora mitochondria. J. Struct. Biol. 129, 48–56

62 Bohm, J. et al. Toward detecting and identifyingmacromolecules in a cellular context: Templatematching applied to electron tomograms. Proc. Natl.Acad. Sci. U. S. A. (in press)