Single Particle Reconstructions of the Transferrin– Transferrin

doi:10.1016/j.jmb.2005.11.021 J. Mol. Biol. (2006) 355, 1048–1065

Single Particle Reconstructions of the Transferrin–Transferrin Receptor Complex Obtained with DifferentSpecimen Preparation Techniques

Yifan Cheng1, Elmar Wolf1, Mykol Larvie2, Olga Zak3, Philip Aisen3

Nikolaus Grigorieff4, Stephen C. Harrison2 and Thomas Walz1*

1Department of Cell BiologyHarvard Medical School, 240Longwood Avenue, BostonMA 02115, USA

2Howard Hughes MedicalInstitute, Children’s Hospitaland Department of BiologicalChemistry and MolecularPharmacology, HarvardMedical School, 320 LongwoodAvenue, Boston, MA 02115USA

3Department of Physiologyand Biophysics, Albert EinsteinCollege of Medicine, 1300Morris Park Avenue, BronxNY 10461, USA

4Howard Hughes MedicalInstitute, Rosenstiel BasicMedical Sciences ResearchCenter, Department of Bio-chemistry, Brandeis University415 South Street, WalthamMA 02454, USA

0022-2836/$ - see front matter q 2005 E

Abbreviations used: Tf, transferrireceptor; EM, electron microscopy;function; SNR, signal-to-noise.

E-mail address of the [email protected]

The outcome of three-dimensional (3D) reconstructions in single particleelectron microscopy (EM) depends on a number of parameters. We haveused the well-characterized structure of the transferrin (Tf)–transferrinreceptor (TfR) complex to study how specimen preparation techniquesinfluence the outcome of single particle EM reconstructions. The Tf–TfRcomplex is small (290 kDa) and of low symmetry (2-fold). Angularreconstitution from images of vitrified specimens does not reliablyconverge on the correct structure. Random conical tilt reconstructionsfrom negatively stained specimens are reliable, but show variable degreesof artifacts depending on the negative staining protocol. Alignment of classaverages from vitrified specimens to a 3D negative stain reference modelusing FREALIGN largely eliminated artifacts in the resulting 3D maps, butnot completely. Our results stress the need for critical evaluation ofstructures determined by single particle EM.

q 2005 Elsevier Ltd. All rights reserved.

Keywords: transferrin receptor–transferrin complex; specimen preparation;angular reconstitution; random conical tilt; cryo-negative staining
*Corresponding author
Introduction

Single particle electron microscopy (EM) hasbecome a standard approach for determining thethree-dimensional (3D) structure of biologicalmolecules at intermediate resolution (20–7 A).Single particle EM has contributed to elucidatingstructures of large macromolecular complexes

lsevier Ltd. All rights reserve

n; TfR, transferrinCTF, contrast transfer

ing author:

several MDa in size, such as viruses,1,2 theribosome3,4 and the spliceosome,5,6 as well as todetermining structures of molecules as small as afew hundred kDa, such as the spliceosomal U1small nuclear ribonucleoprotein particle,7 a-latro-toxin oligomers,8 and the transferrin (Tf)–transfer-rin receptor (TfR) complex.9 A problem in singleparticle EM is that it is often difficult to decidewhether a 3D reconstruction is correct, if no otherstructural information is available. The quality andthus the reliability of a 3D reconstruction are oftenjudged on the basis of two criteria. First, a plot of theangular distribution of the particle images revealswhether the data set contains all the views that are

d.

3D Reconstructions of the Tf–TfR Complex 1049

necessary to completely define the structure.Second, comparison of projections from the recon-structed 3D map with the class averages or the rawparticle images can verify that the 3D structure isconsistent with the projection data. The presence ofall required views thus validates that the structureof the molecule is completely defined, whereas agood match between the density map and theprojection data confirms that the 3D reconstructionis consistent with the raw data. While it is essentialto assess a 3D reconstruction by both of thesecriteria, neither one addresses the question ofwhether the final reconstruction is indeed a correctrepresentation of the imaged molecule. It is there-fore important to understand the factors thatinfluence and possibly distort the outcome of a 3Dreconstruction. Here, we have undertaken an arrayof experiments to study the influence of thespecimen preparation technique on the quality ofthe single particle reconstructions that can beobtained.

Two approaches are currently used in singleparticle EM, as opposed to electron tomography, tocalculate a 3D reconstruction from projectionimages: random conical tilt10 and angular recon-stitution.11 In the random conical tilt method, thespecimen is imaged twice, first at a high tilt angleand then untilted. The particles from the images ofthe untilted specimen are translationally androtationally aligned to each other to determine thex, y coordinates of the center and the in-planerotations of all the particles. Together with the tiltangle chosen for imaging the specimen, all theorientational parameters needed to reconstruct a 3Dmap are thus unambiguously defined. For thisapproach to be practical, the specimen shouldadsorb to the carbon support in one or only a feworientations. This is often the case for negativelystained specimens where the particles areembedded in a layer of heavy-metal salts.12

Preferred orientations are particularly beneficialwhen working with heterogeneous specimens, asthey allow separation of particles in differentconformations by classification (recently illus-trated13). The random conical tilt 3D reconstructionapproach applied to images of negatively stainedspecimens thus allows one to calculate 3D recon-structions of molecules in different conformationsfrom images of the same EM grid. Negative stainingis known, however, to introduce specimen prep-aration artifacts. For example, incomplete stainembedding renders invisible the part of the proteinthat protrudes from the stain layer and drying of thegrid usually results in flattening of the specimen.

By contrast, vitrification, the embedding of asample in a thin layer of vitrified ice,14 introducesvirtually no artifacts and is considered the bestspecimen preparation technique. Furthermore,unlike negative staining, sample vitrification doesnot limit the resolution that can be achieved.Vitrified particles usually adopt more or lessrandomly distributed orientations in theamorphous ice layer, which can be exploited

to calculate an initial 3D reconstruction using theangular reconstitution approach. Five parameters,the x, y coordinates of the center and the three Eulerangles, have to be determined for each particle tocalculate an initial density map. Euler angles for theparticles are determined using the common-linesmethod.15 The contrast of images taken fromvitrified specimens is very low, because of thesmall difference in electron scattering between thevitrified ice layer and the embedded molecules.Therefore, the particles are first grouped into anappropriate number of classes to produce averageswith an improved signal-to-noise ratio (SNR). Theseare then used for Euler angle assignment. Angularreconstitution using images of vitrified specimenssuffers from two problems. First, vitrification isoften not suitable for heterogeneous specimens,because it is not always possible to decide whethertwo dissimilar images are indeed images of twodifferent particles, e.g. particles in different confor-mations, or are images of the same particle but indifferent orientations. Second, angular reconstitu-tion attempts to solve a problem that underunfavorable conditions can have more than onesolution. A 3D structure will always produce thesame set of projection images, but if certainimportant views are missing, leaving the 3Dstructure partially undefined, a set of projectionimages can produce different 3D reconstructionsthat are all consistent with the projection data. Oncean initial model has been produced, by eithermethod, the 3D reconstruction is improved byrefinement of the orientational parameters of theparticle images. Although different strategies havebeen implemented in different program packages,e.g. IMAGIC,16 SPIDER,17 EMAN,18 and FREA-LIGN,19 they all rely on iterative cycles of refiningthe alignment of the projection images to a reference3D model.

We have previously used single particle cryo-electron microscopy to determine the structure ofthe human Tf–TRf complex, a 2-fold symmetriccomplex with a molecular mass of 290 kDa, to anominal resolution of 7.5 A.9 By docking the crystalstructures of Tf and TRf into the density map, wecould build an atomic model for this complex. In thework reported here, we have used the structurallywell characterized Tf–TfR complex to systema-tically assess the influence of the specimen prep-aration technique on the outcome of a 3Dreconstruction obtained by single particle electronmicroscopy.

Results

Vitrified Tf–TfR complex

Recombinant human Tf–TfR complex wasapplied to holey carbon film, blotted with filterpaper and vitrified by plunging into liquid ethane.Individual complexes could clearly be identified inlow-dose images of such preparations (Figure 1(a)),

Figure 1. Vitrified Tf–TfR complex. (a) Low-dose image of vitrified Tf–TfR complexes prepared using holey carbonfilm. (b) Gallery of representative class averages displaying various views of the Tf–TfR complex. (c) Plot of the Eulerangle distribution showing that the complex has a tendency to adopt preferred orientations in the vitrified ice layer.(d) Views of the best initial 3D reconstruction obtained by the angular reconstitution approach implemented in theIMAGIC software. (e) The same views as in (d) for the atomic model of the Tf–TfR complex. The lines in the first twopanels indicate the 2-fold axis. (f) Low-dose image of Tf–TfR complexes vitrified on a continuous carbon film. (g) Galleryof representative class averages revealing that Tf–TfR complexes vitrified on a continuous carbon film do not adoptpreferred orientations. (h) Plot of the Euler angle distribution showing that the particles adsorb to the carbon film inrandom and almost completely uniformly distributed orientations. The scale bars in (a) and (f) represent 50 nm and theindividual panels in (b) and (g) have a side length of 27 nm.

1050 3D Reconstructions of the Tf–TfR Complex

and we interactively selected about 36,000 particlesfor digital image processing. We used the IMAGICsoftware package16 to classify the particle imagesinto 500 classes by several cycles of multi-referencealignment (MRA) and multi-variance statisticalanalysis (MSA) followed by classification(Figure 1(b)). When we used the angular reconstitu-tion algorithm implemented in IMAGIC to generatean initial model (see Materials and Methods for

details), we found that different attempts resulted invery different density maps. The 3D mapsdepended not only on the class averages weselected, but also on the order in which we introducedthem into the reconstruction procedure. This problemhas been recognized before and was addressed by analgorithm that simultaneously determines theorientations for a set of projections.20 However,this alternative approach did not completely resolve


the problem, as the density maps still depended onthe number and sets of projection averages that wereincluded in the reconstruction (data not shown). Wewould therefore not have been able to decide whichdensity map to choose for further structural refine-ment if no additional structural information had beenavailable. Because the crystal structures of Tf21 andTfR22 were available, we were able to identify the 3Dmap produced by angular reconstitution that wasmost consistent with these crystal structures,although even this map (Figure 1(d)) looked onlyremotely similar to the final atomic model weproduced for the Tf–TfR complex9 (Figure 1(e)).Using FREALIGN19 to refine the orientationalparameters of the particles against this initial model,we obtained the previously published density map ata nominal resolution of 7.5 A.9 A plot of the Eulerangles showed that views were distributed more orless across the entire orientational space, althoughthere was a clear preference for some of the particleorientations (Figure 1(c)).

We were concerned by the different initialmodels produced by the angular reconstitutionapproaches. We hoped to overcome this problemby calculating a 3D reconstruction using randomconical tilt, because here the orientational par-ameters of the particles are uniquely defined. Forthis approach to be practical, the particles shoulddisplay preferred views. Therefore, in an attemptto induce preferred orientations, we vitrified Tf–TfR complexes that had been adsorbed to acontinuous carbon film. Low-dose images takenfrom such preparations looked very similar tothose taken from samples on holey carbon film,although the image contrast was slightly weakerdue to the noise introduced by the carbonsupport (Figure 1(f)). We interactively selected21,719 particles and classified them into 200classes. The resulting class averages showedmany different views of the Tf–TfR complex,demonstrating that adsorption to continuouscarbon film did not introduce preferred orien-tations of the complexes (Figure 1(g)). Since thispreparation was not suitable for obtaining a 3Dreconstruction by random conical tilt, we usedFREALIGN to align the particle images to ourpreviously determined 3D map of the Tf–TfRcomplex. A plot of the Euler angles (Figure 1(h))revealed that the orientations adopted bycomplexes vitrified on a continuous carbon filmwere also random and even more uniformlydistributed than those of complexes vitrified inholes of holey carbon film (Figure 1(c)).

We finally collected tilt pairs of vitrified Tf–TfRcomplexes in holey carbon film, assuming that ifwe collected a large number of images, a suitablenumber of the molecules would be in a particularorientation. We found, however, that the Tf–TfRcomplex is too small to be clearly seen in imagesof tilted specimens due to the thicker ice layer(data not shown), making it impossible todetermine the structure of the vitrified Tf–TfRcomplex by random conical tilt.

Negatively stained Tf–TfR complex

Vitrification of Tf–TfR complex using holey orcontinuous carbon film did not show clear pre-ferred orientations and was therefore not suitablefor calculating 3D reconstructions by randomconical tilt. Previous EM experiments showed,however, that preferred adsorption of the Tf–TfRcomplex to the carbon support could be induced bynegative staining.13

Conventional negative staining with uranyl formate

A Tf–TfR complex sample was adsorbed to aglow-discharged carbon film and negativelystained with uranyl formate as described before.13

The specimen was imaged at tilt angles of 608 and 08(Figure 2(a) and (b)). Since in the conventionalnegative staining technique the particles are notembedded in a continuous layer of stain, images of608 tilted samples show dark shadows below themolecules resulting from stain accumulation (stainclouds) surrounding individual particles(Figure 2(b)). We interactively selected 8410 pairsof particles from 80 image pairs using the displayprogram WEB associated with the SPIDER softwarepackage.17 Classification calculations showed thatmost of the particles from images of the untiltedsample fell into two classes. A majority of theparticles adsorbed to the grid presented a side view(Figure 2(c), panel 1), while most of the remainingparticles presented a top view (Figure 2(c), panel 2).

We combined the images of all classes thatshowed the same view of the complex and a crispfine structure, and we used the combined particlesfrom the images of the tilted sample to calculate 3Dreconstructions. We first calculated a 3D map of theparticles presenting the side view (Figure 2(c),panel 1). The unsymmetrized map clearly resolveda density representing the TfR as well as densitiesrepresenting both the N and the C-lobe of the twobound Tf molecules (Figure 2(d)). The map alsorevealed distortions. The first view of the recon-struction shows that one of the Tf C-lobes (markedby an asterisk in Figure 2(d)) is clearly smaller thanthe other C-lobe and the two N-lobes. Furthermore,the third and fourth views of the reconstructionshow the left side of the complex to be flat. This flatsurface most likely reflects a deformation of thecomplex resulting from its interaction with thecarbon film. Since the Tf–TfR complex is 2-foldsymmetric, we applied this symmetry to thereconstruction and then fit the TfR and Tf crystalstructures into the resulting density map(Figure 2(e)). The 2-fold symmetrization of themap removed the distortions mentioned above,rendering all four Tf lobes of equal size andeliminating the flat surface where the moleculeinteracts with the carbon film. The fit of crystalstructures showed, however, that the density for theTfR was still substantially deformed. Rather thanrevealing the characteristic butterfly shape of theTfR dimer, the second view of the symmetrized

Figure 2. Tf–TfR complex prepared by conventional negative staining. (a) and (b) Low-dose images of negativelystained preparations recorded at tilt angles of (a) 08 and (b) 608. The image of the tilted specimen shows the characteristicshadows below the particles. (c) Class averages showing the two preferred orientations in which the Tf–TfR complexadsorbs to the carbon film; panel 1, side view; panel 2, top view. (d) Views of the unsymmetrized random conical tiltreconstruction obtained with the side view particles (class 1 in (c)) showing a smaller density for one of the Tf C-lobesand loss of the characteristic butterfly shape of the receptor. (e) Same density map as in (d) after 2-fold symmetrization.Fitting of the Tf and TfR crystal structures into the density map reveals that the 3D reconstruction is flattened,particularly the density representing the TfR. (f) 2-fold symmetrized 3D reconstruction using the top view particles (class2 in (c)) revealing substantial flattening and a lack of distinct structural features. The scale bar in (a) represents 50 nm andthe individual panels in (c) have a side length of 27 nm.


reconstruction in Figure 2(e) shows a featurelessdensity with no indication of the bipartite organi-zation of the TfR dimer. The density for the TfR isalso severely affected by flattening. Rather than thewidth of 13.7 nm in the crystal structure, the densitymap representing the TfR has a width of only7.5 nm, a flattening of about 45%.

We also calculated a 3D reconstruction using theparticles that had adsorbed to the grid presenting atop view (Figure 2(c), panel 2). While flatteningmostly affects the density representing TfR in mapsreconstructed from “side view particles”(Figure 2(e)), it affects the entire complex in the 3Dreconstruction of “top view particles” (Figure 2(f)). Asseen in the first two views of the reconstruction in

Figure 2(f), the atomic model protrudes from the 3Dmap due to a substantial flattening of w55%.Probably because of this flattening, the unsymme-trized map (not shown) as well as the 2-foldsymmetrized map (Figure 2(f)) show very fewfeatures of the Tf–TfR complex. All the densities inthe symmetrized map are fused, resolving neither theseparation between TfR and Tf nor the cleft betweenthe Tf N and C-lobes.

Negative staining with uranyl formate using a carbonfilm sandwich

Incomplete stain embedding renders invisibleany protein domains protruding from the stain


layer. Since the images taken from the 608 tiltedsamples showed prominent stain clouds(Figure 2(b)), incomplete stain embedding provideda potential explanation for the flattening seen in the3D reconstructions (Figure 2(d)–(f)). To address thisproblem, we employed the carbon sandwichtechnique, previously used to visualize, forexample, the ribosome.23 The Tf–TfR complex wasadsorbed to a single layer of carbon film andnegatively stained with uranyl formate. Anotherlayer of carbon film was then deposited on thespecimen, so that the Tf–TfR complexes were nowsandwiched between two layers of carbon andembedded in a continuous layer of stain. The detailsof this procedure have been described before.13 Wewill refer to this specimen preparation techniquefrom here on as the carbon sandwich technique.

The appearance of Tf–TfR complexes prepared bythe carbon sandwich technique varied greatly indifferent areas of the EM grid. In some areas, theparticles were still surrounded by stain clouds,suggesting a thin layer of stain accompanied byincomplete stain embedding of the molecules.Occasionally, we found areas where the moleculeswere embedded in a thick layer of stain, showingfine structure and no stain clouds even at a tilt of 608(Figure 3(a) and (b)). In other areas the particlesappeared larger and showed almost no finestructure, indicating that the molecules becamesquashed between the two carbon layers upondrying, resulting in massive deformation of themolecules. We used image areas where the particleswere mostly unsquashed and did not show stainclouds at a tilt of 608 to collect 608/08 image pairs.We interactively selected 14,820 pairs of particles.The particles were classified as before, againrevealing that the complexes adopt two predomi-nant orientations on the carbon film (Figure 3(c)),but the size of the particles in the class averagesvaried. The size of the smallest particles(Figure 3(c), panels 1 and 3) corresponded wellwith that of particles in the class averages of theconventionally stained Tf–TfR complexes(Figure 2(c)). Some class averages showed particleswith the same features but a larger size (examplesare shown in Figure 3(c), panels 2 and 4). Finally,many class averages showed even larger particlesthat no longer displayed any discernible structuralfeatures of a Tf–TfR complex. An example for anaverage of such squashed particles is shown inpanel 5 of Figure 3(c).

The images that produced the smallest classaverages were first used to calculate 3D reconstruc-tions. The unsymmetrized 3D reconstruction(Figure 3(d)) obtained with the particles thatproduced class average 1 (Figure 3c, panel 1)looks substantially more like the undistortedcomplex than does the corresponding reconstruc-tion from the Tf–TfR particles prepared by conven-tional negative staining (Figure 2(d)). All four Tflobes are represented by densities of almost thesame size, and it is not immediately obvious fromthe 3D map with which side the particle has

adsorbed to the carbon film. The density represent-ing the TfR dimer also shows more fine structurethan before and indicates a bipartite organization.The left side of the receptor dimer in the secondpanel of Figure 3(d) appears, however, to have sliddown compared to the right side, obscuring thepresence of a 2-fold symmetry in the receptor. Afterapplying the 2-fold symmetry, the density mapreveals all the characteristic features of the Tf–TfRcomplex (Figure 3(e)), including, to some degree,the butterfly shape of the receptor (Figure 3(e),panel 2), but the carbon sandwich technique doesnot overcome the flattening problem, as the densitymap is flattened by about 38%.

When we used the top view particles (Figure 3(c),panel 3), the resulting 2-fold symmetrized densitymap (Figure 3(f)) also looked more like theundistorted complex than the comparable recon-struction from particles prepared by conventionalnegative staining (Figure 2(f)). As before, theseparticles suffered more severe flattening (w45%)than the side view particles (w38%), and thefeatures of the Tf–TfR complex were less wellresolved. Nevertheless, unlike the reconstructionfrom particles prepared by conventional negativestaining, there is some separation between thedensities for TfR and Tf, and there is someindication for the two lobes in the bound Tfmolecules.

Finally, we also calculated a 3D reconstruction forone of the classes showing severely deformedTf–TfR complexes. Like the projection average(Figure 3(c), panel 2), the unsymmetrized 3Dreconstruction shows a more extended molecule(Figure 3(g)). In addition, the reconstruction isflatter than the one in Figure 3(d) (6.5 nm insteadof 8.5 nm). The Tf lobes appear to be further apartfrom each other and further removed from the TfR,and the two TfR halves seem to have slid evenfurther apart from each other, removing anyindication of dimeric organization (Figure 3(g),panel 2). The whole appearance of this 3Dreconstruction is consistent with a “spread-out”Tf–TfR complex, perhaps the result of forces exertedby the two carbon layers that approach each otheras the specimen dries.

Cryo-negative staining with uranyl formate usinga carbon sandwich and 5% glycerol

The carbon sandwich technique improved thestain embedding and yielded 3D maps with morefine structure, but the 3D reconstructions stillsuffered from severe flattening and deformations.To address these remaining problems, we added 5%glycerol to the sample prior to staining with uranylformate. We then added the second carbon film andfroze the sample by plunging it into liquid nitrogenwith the glycerol serving to increase the viscosity ofthe buffer and as cryo-protectant. This preparationtechnique should conserve the complete stainembedding provided by the carbon sandwich andin addition prevent drying artifacts. The detailed

Figure 3. Tf–TfR complex prepared by the carbon sandwich technique. (a) and (b) Low-dose images of carbonsandwiched preparations recorded at tilt angles of (a) 08 and (b) 608. The image of the tilted specimen does not showshadows below the particles, which is characteristic for the conventional negative staining technique. (c) Representativeclass averages; panel 1, side view; panel 2, side view of flattened particles; panel 3, top view; panel 4, top view of flattenedparticles; panel 5, average of badly distorted particles showing no recognizable features of a Tf–TfR complex. (d) Viewsof the unsymmetrized random conical tilt reconstruction obtained with the side view particles (class 1 in (c)). The densityrepresenting the receptor is asymmetric and it appears as if the left half of the receptor has slid down with respect to itsright half. (e) Same density map as in (d) after 2-fold symmetrization restoring, to some degree, the butterfly shape of thereceptor. Fitting of the Tf and TfR crystal structures into the density map reveals that the 3D reconstruction is stillflattened. (f) The 2-fold symmetrized 3D reconstruction using the top view particles (class 3 in (c)) revealing substantialflattening, but more structural features than the corresponding reconstruction using complexes prepared by theconventional negative stain technique. (g) Unsymmetrized 3D reconstruction obtained with flattened side view particles(class 2 in (c)) revealing a spread-out morphology of the complex. The scale bar in (a) represents 50 nm and theindividual panels in (c) have a side length of 27 nm.


procedure for this sample preparation technique,which we will simply call “cryo-negative staining”(a modified procedure from the one introduced byStark and co-workers24 and different from the

technique introduced by Dubochet and co-workers25), has been described before.13

Cryo-negatively stained samples were imagedat liquid nitrogen temperature. As in the case of


the carbon sandwiched samples, the thickness ofthe stain layer varied in different areas of thegrid, and Tf–TfR complexes in areas with a thinlayer of stain showed the typical stain clouds. Tiltpairs (508 and 08) were thus collected only fromareas with a thick layer of stain. Despite strictlow-dose procedures, the images often showedwhite bubbles, particularly in areas with a thicklayer of stain. The occurrence of such beam-induced bubbles is a sign of radiation damageand, accordingly, particles in images with bubbleswere often of poor quality, having a round shapeand no obvious fine structure. We observedthe bubbles more frequently in the images ofthe untilted sample, which were taken after thespecimen was already imaged at a 508 tilt. Sinceparticles from images of untilted specimens thatshowed bubbles aligned poorly to each other, thecorresponding particles from the images of the508 tilted sample could not be used. In the end,only a small percentage of the imaged particleswere suitable for calculating 3D reconstructions.

Figure 4(a) and (b) show the two images of agood tilt pair. The particles are embedded in athick layer of stain, and neither image is visiblyaffected by radiation damage, so that the imageof the untilted sample still shows clear structuralfeatures of the Tf–TfR complex (Figure 4(a)). Weinteractively selected a total of 23,142 particlepairs from 48 image pairs and classified theparticles from the images of the untilted sample.Figure 4(c) shows representative class averages.Class averages 1 to 3 show side views of theTf–TfR complex. The particles represented byaverage 1 have the expected size and theexpected features of the complex, while thoserepresented by average 2 are larger, indicatingthat they are squashed. Average 3 also representssquashed particles, but in addition this averagedoes not show any fine structure of the complex,indicating that these molecules sustained furtherdamage, possibly related to the radiation-inducedbubbles. Class averages 4 and 5 show top viewsof the Tf–TfR complexes. Average 4 representsparticles of the correct size, whereas average 5contains severely squashed particles. Finally, classaverage 6 represents molecules that are so badlyaffected by flattening and beam damage that theyno longer show any distinct structural features.

The unsymmetrized 3D reconstructioncorresponding to class average 1 shows cleardensities for all four Tf lobes and the TfR dimer.The flattening is negligible (Figure 4(d)). More-over, the second view of the reconstruction showsthat the density representing TfR reflects thebutterfly shape of the receptor dimer, although itis somewhat deformed. After 2-fold symmetriza-tion, the density map is consistent with the TfRand Tf crystal structures, as seen by their good fitinto the map (Figure 4(e)). It is noteworthy thatthis reconstruction was obtained with only 4190particles, a sixth of the entire data set. The 3Dreconstruction calculated with the top view

particles represented by class average 4 showssignificantly less flattening (w34%) (Figure 4(f))than the corresponding reconstructions fromparticles prepared by conventional negativestaining (Figure 2(f)) or the carbon sandwichtechnique (Figure 3(f)). The density map shows,however, fewer structural features than the mapobtained from carbon sandwiched top viewparticles (Figure 3(f)). For comparison, we alsocalculated a 3D reconstruction of a class of largerparticles (Figure 4(c), average 2). While thedensity map (Figure 4(g)) shows the same“spread-out” appearance already seen in thecorresponding reconstruction from carbon sand-wiched particles (Figure 3(g)), the molecule doesnot seem to be significantly flattened.

Negative staining with ammonium molybdatein 2% glucose

A mixture of ammonium molybdate andglucose has previously been used to visualizekeyhole limpet hemocyanin type 1, which pro-duced a density map at 15 A resolution.26

Glucose embedding is commonly used forpreparing two-dimensional crystals,27,28 butammonium molybdate has to be added forwork on single particles to increase the contrastbetween the molecules and the embeddingmedium. The Tf–TfR complex sample wasmixed at a 1:1 ratio with a solution of 2% (w/w)glucose and 1% (w/w) ammonium molybdate andapplied to a glow-discharged grid before freezing.The mixing of the sample prior to its adsorption tothe grid ensured that the Tf–TfR complexes werecompletely embedded in the stain solution. Theimages taken from such preparations (Figure 5(a))show significantly less contrast than those takenfrom specimens stained with uranyl formate oreven those from vitrified specimens (e.g.Figure 1(a)).

A total of 17,107 particles were interactivelyselected from 39 micrographs and subjected tomultiple rounds of MRA and classification usingthe IMAGIC software package. The resultingclass averages revealed that the Tf–TfR complexesassumed many different orientations (Figure 5(b)),similar to complexes embedded in a vitrified icelayer. A 3D reconstruction using the randomconical tilt approach was thus impractical. More-over, despite significant effort, we could notproduce a reasonable initial 3D model using theangular reconstitution approach. We thereforeused the 2-fold symmetrized 3D reconstructionobtained with vitrified Tf–TfR complex9 as aninitial model, to which we aligned the particleimages with FREALIGN. A plot of the Eulerangles showed that the particles indeed adoptedrandom and almost uniformly distributed orien-tations (Figure 5(c)). The Fourier shell correlation(FSC) curve of the final 3D reconstruction usingthe FSCZ0.15 cut-off criterion29 suggested aresolution of 17 A (data not shown), better than

Figure 4. Tf–TfR complex prepared by cryo-negative staining. (a) and (b) Low-dose images of cryo-negatively stainedspecimens recorded at tilt angles of (a) 08 and (b) 508. (c) Representative class averages; panel 1, side view; panel 2, sideview of flattened particles showing clear densities for TfR and Tf; panel 3, side view of flattened particles showing littlestructural features; panel 4, top view; panel 5, top view of flattened particles; panel 6, average of badly distorted particlesshowing no recognizable features of a Tf–TfR complex. (d) Views of the unsymmetrized random conical tiltreconstruction obtained with the side view particles (class 1 in (c)). Although the density representing the receptor issomewhat asymmetric, the map shows little flattening. (e) Same density map as in (d) after 2-fold symmetrization withthe Tf and TfR crystal structures docked into the density map. The good fit of the crystal structures into the density mapindicates that the particles suffered little distortion or flattening. (f) A 2-fold symmetrized 3D reconstruction using thetop view particles (class 4 in (c)). Although the density map shows less flattening than the corresponding reconstructionsusing particles prepared by conventional negative staining and the carbon sandwich technique, the densitiesrepresenting Tf and TfR are still poorly resolved. (g) Unsymmetrized 3D reconstruction obtained with flattened sideview particles (class 2 in (c)). The density map shows a spread-out particle, but little flattening. The scale bar in (a)represents 50 nm and the individual panels in (c) have a side length of 27 nm.


any reconstruction obtained with uranyl formatestaining, and the map was filtered to thisresolution (Figure 5(d)). While the density mapreveals the correct overall shape of the complex

with no major distortions, docking of the TfR andTf crystal structures showed that the crystalstructures protrude from the 3D map in variousplaces.

Figure 5. Tf–TfR complex prepared with a mixture of ammonium molybdate and glucose. (a) Low-dose image ofvitrified Tf–TfR complexes showing the poor image contrast obtained with this preparation technique. (b) Gallery ofrepresentative class averages displaying various views of the Tf–TfR complex. (c) Plot of the Euler angle distributionshowing that the complex adopts completely random orientations. (d) Views of the 3D reconstruction obtained by usingFREALIGN to align the class averages to the 3D map obtained with images of vitrified samples. The scale bar in (a)represents 50 nm and the individual panels in (b) have a side length of 27 nm.


Use of FREALIGN to align images of vitrifiedTf–TfR complexes to 3D maps obtained byrandom conical tilt

The 2-fold symmetrized random conical tiltreconstruction obtained with the cryo-negativelystained particles was a very good representation ofthe structure of the Tf–TfR complex, with virtuallyno signs of distortion or flattening. We thereforetested whether we could use this map to align theparticle images we recorded from the vitrifiedTf–TfR complex sample. For this purpose we usedFREALIGN,19 taking full advantage of the 2-foldsymmetry of the Tf–TfR complex for both alignmentand 3D reconstruction. With FREALIGN we had thechoice of either aligning the individual particleimages or the class averages to the reference model.Since images taken from vitrified particles are verynoisy and have low contrast, we chose to use theclass averages whose alignment to the referencemodel should be less sensitive to spurious featuresin either the model or the images themselves due totheir higher SNR. The individual images of thevitrified Tf–TfR complexes were therefore band-pass filtered to include spatial frequencies between1/40 AK1 and 1/170 AK1, including only databefore the first node of the CTF, and grouped into500 classes. The large number of classes wasnecessary to ensure that the class averages alsorepresented particles adopting less abundant orien-tations. In FREALIGN, the class averages arealigned directly to the reference model. The Eulerangles thus determined were then used to calculate

a new 3D map from the class averages, and this 3Dmap served, in turn, as a reference model foriterative refinement of the orientational parametersfor the class averages. The density map after 15cycles (Figure 6(b)) was very similar to the 40 Adensity map calculated from the atomic model ofthe complex (Figure 6(a)).

Applying 2-fold symmetry to the 3D reconstruc-tion of the Tf–TfR complex was essential toeliminate distortions in the density map obtainedwith the cryo-negatively stained sample (compareFigure 4(d) and (e)). We therefore decided to testwhat would happen in the case of a molecule withno inherent symmetry. Vitrification does not intro-duce significant distortions in the molecules,whereas cryo-negative staining does. We thereforewondered whether alignment of the class averagesfrom the undistorted vitrified particles wouldrectify the distortions seen in the reference modelobtained from the cryo-negatively stained speci-mens. Alternatively, it was possible that thedistortions of the reference model would imposedistortions on the final density map obtained byaligning the class averages from the vitrifiedparticles to the distorted reference model. Tosimulate an asymmetric molecule, we used as thereference model the density map of cryo-negativelystained Tf–TfR complex produced by randomconical tilt without applying 2-fold symmetry(Figure 4(d)). Alignment of the class averagesfrom images of the vitrified complexes and 3Dreconstruction using FREALIGN was also per-formed without enforcing 2-fold symmetry.

Figure 6. Use of random conical tilt reconstructions for the alignment of class averages obtained from images ofvitrified particles. (a) Density map obtained by resolution filtering the atomic model of the Tf–TfR complex to 40 A.(b)–(f) Density maps obtained with FREALIGN. Euler angles were assigned to the class averages calculated from theimages of vitrified complexes by directly aligning them to reference maps. The reference maps used were: (b) 2-foldsymmetrized 3D reconstruction of the side view particles in cryo-negatively stained preparations; (c) unsymmetrized 3Dreconstruction of the side view particles in cryo-negatively stained preparations; (d) 2-fold symmetrized 3Dreconstruction of the top view particles in cryo-negatively stained preparations; (e) 2-fold symmetrized 3Dreconstruction of the side view particles in samples prepared by conventional negative staining; (f) unsymmetrized3D reconstruction of the side view particles in samples prepared by conventional negative staining.


The resulting density map showed a reduction ofthe deformations seen in the 3D reconstructionof the cryo-negatively stained reconstruction, butdid not fully restore the 2-fold symmetry of thecomplex (compare Figure 6(c) with (b)).

We also sought to investigate whether FREA-LIGN could restore a meaningful density map fromthe class averages of the vitrified specimen if thereference model was substantially distorted.We first used the 3D reconstruction obtained with

the cryo-negatively stained top view particles,which showed not only substantial flattening butalso a lack of distinct structural features(Figure 4(f)). Alignment of the class averages fromthe vitrified specimen to this reference modelresulted in a density map that only slightlyresembled the structure of the Tf–TfR complex,even when 2-fold symmetry was enforced(Figure 6(d)). We then used random conical tiltreconstructions obtained from particles negatively

Ta

ble

1.

Co

mp

aris

on

of

spec

imen

pre

par

atio

nte

chn

iqu

esu

sin

gth

eT

f–T

fRco

mp

lex

Co

nv

enti

on

aln

egat

ive

stai

nin

gC

arb

on

san

dw

ich

Cry

o-n

egat

ive

stai

nin

gA

mm

on

ium

mo

lyb

dat

eV

itri

fied

inh

ole

sV

itri

fied

on

carb

on

Ori

enta

tio

no

fp

arti

cles

Pre

ferr

edP

refe

rred

Pre

ferr

edR

and

om

Ran

do

m(s

lig

htl

yp

re-

ferr

eda)

Ran

do

mb

3Dre

con

stru

ctio

nap

pro

ach

Ran

do

mco

nic

alti

ltR

and

om

con

ical

tilt

Ran

do

mco

nic

alti

ltA

ng

ula

rre

con

stit

uti

on

cA

ng

ula

rre

con

stit

uti

on

cA

ng

ula

rre

con

stit

uti

on

c

Sp

ecim

enfl

atte

nin

g45

%(s

ide)

38%

(sid

e)0%

(sid

e)“N

ofl

atte

nin

g”d

No

flat

ten

ing

No

flat

ten

ing

55%

(to

p)

45%

(to

p)

34%

(to

p)

Sp

ecim

end

isto

rtio

ns

Sev

ere

dis

tort

ion

sS

ever

ed

isto

rtio

ns

Min

imal

dis

tort

ion

s“N

od

isto

rtio

ns”

dN

od

isto

rtio

ns

No

dis

tort

ion

sA

dv

anta

ges

Hig

hco

ntr

ast,

easy

dat

aco

llec

tio

nH

igh

con

tras

t,ea

syd

ata

coll

ecti

on

Lea

stfl

atte

nin

g,

leas

td

isto

rtio

nsd

“No

flat

ten

ing

”d“n

od

isto

rtio

ns”

Per

fect

spec

imen

pre

s-er

vat

ion

Per

fect

spec

imen

pre

s-er

vat

ion

Dra

wb

ack

sS

ever

efl

atte

nin

g,

sev

ere

dis

tort

ion

sS

ever

efl

atte

nin

g,

sev

ere

dis

tort

ion

sD

iffi

cult

dat

aco

llec

tio

nN

ot

go

od

for

ran

do

mco

nic

alti

ltc,

eL

ow

con

tras

t,n

ot

go

od

for

ran

do

mco

nic

alti

ltc,

eL

ow

con

tras

t,n

ot

go

od

for

ran

do

mco

nic

alti

ltc,

e

aP

arti

cles

vit

rifi

edo

nh

ole

yca

rbo

nfi

lms

usu

ally

ado

pt

ran

do

mo

rien

tati

on

s,b

ut

som

eo

rien

tati

on

of

the

par

ticl

esca

no

ccu

rat

the

air/

wat

erin

terf

ace.

bT

his

resu

ltw

ith

the

Tf–

TfR

com

ple

xis

no

tty

pic

al,

sin

cev

itri

fica

tio

no

nco

nti

nu

ou

sca

rbo

nfi

lms

usu

ally

ind

uce

sp

refe

rred

ori

enta

tio

ns.

cS

amp

les

wit

hp

arti

cles

inra

nd

om

ori

enta

tio

ns

can

also

be

use

dfo

rra

nd

om

con

ical

tilt

reco

nst

ruct

ion

s,b

ut

this

req

uir

esth

atv

ery

larg

ed

ata

sets

are

coll

ecte

d.

dT

he

par

ticl

esp

rep

ared

mig

ht

stil

lb

efl

atte

ned

and

dis

tort

ed,

bu

tth

isis

no

tv

isib

lein

3Dre

con

stru

ctio

ns

calc

ula

ted

by

ang

ula

rre

con

stit

uti

on

.e

“No

tg

oo

dfo

rra

nd

om

con

ical

tilt

”is

list

edh

ere

asa

dra

wb

ack

,b

ecau

seo

fth

eu

nce

rtai

nti

esan

gu

lar

reco

nst

itu

tio

nca

nin

tro

du

cein

the

3Dre

con

stru

ctio

n.


stained in the conventional way as referencemodels. Using the 2-fold symmetrized 3D map ofthe side view particles (Figure 2(e)) to align the classaverages of the vitrified specimen produced adensity map in which the flattening of the moleculewas no longer obvious (Figure 6(e)). In addition, thedensity representing the receptor showed a shapemuch more closely reflecting the true structure ofthe receptor domain than in the initial model. Bycontrast, using the 2-fold symmetrized 3D map ofthe top view particles (Figure 2(f)) as the referencemodel resulted in a completely uninterpretable map(data not shown). Finally, we used the unsymme-trized side view particle reconstruction fromparticles negatively stained in the conventionalway as reference model (Figure 2(d)) and also didnot use the symmetry for the refinement. Thisimitation of a distorted asymmetric particleproduced a density map (Figure 6(f)) that wasremarkably similar to the one using the virtuallyunflattened reference model from the cryo-negativestain preparation (Figure 6(c)).

Discussion

All the specimen preparation strategies we testedto calculate a single particle 3D reconstruction forthe Tf–TfR complex resulted in either an ambiguousdensity map or a density map affected to varyingdegrees by preparation artifacts. Our results thushighlight the problem of obtaining a reliable 3Dreconstruction by single particle EM with currentspecimen preparation techniques, drawing atten-tion to the need for careful interpretation of theobtained density maps. While we have used onlyone test specimen in this study, the Tf–TfR complexis a challenging specimen, and because of its“spiky” structure, it is particularly sensitive todeformations. We therefore believe that many ofour findings can be generalized to other specimens,although other specimens may suffer more or lessfrom preparation artifacts depending on their shapeand density distribution (“compactness”). What wehave learned from our 3D reconstructions issummarized in Table 1 and elaborated in thefollowing paragraphs.

Influence of the specimen preparation techniqueon the orientation of the molecules

As expected, vitrified Tf–TfR complexes adoptrandom, more or less uniformly distributed orien-tations in holes of holey carbon films (Figure 1(c)).Not expected was our finding that adsorbing thecomplexes to a continuous carbon film prior tovitrification generated even more randomly dis-tributed orientations (Figure 1(h)). This resultshows that adsorption of molecules to a carbonfilm alone does not guarantee that particles adsorbin preferred orientations. The tendency for pre-ferred orientations of the complexes seen in vitrifiedsamples using holey carbon films may reflect an


alignment of the molecules at the air–water inter-face, which has been observed before.30 This notionis somewhat strengthened by our observation thatTf–TfR complexes embedded in a thicker layer ofvitrified ice showed a reduced tendency forpreferred orientations (data not shown). The moreevenly distributed orientations of the moleculesadsorbed to a continuous carbon film might there-fore be a result of the molecules interacting with thecarbon film, keeping them away from the orientingeffect of the air–water interface.

Consistent with our experience with all theproteins that we have worked with to date, allnegative staining techniques using uranyl formateinduced the Tf–TfR complex to adsorb to the carbonsupport in preferred orientations. In these tech-niques the specimen was adsorbed to the grid,washed with several drops of water, and then eitherdried or frozen with or without prior application ofa second carbon film. Since in the cryo-negativestaining procedure the grid is frozen before it iscompletely dry, complete drying of the grid cannotbe the only reason leading to preferred orientationsof the molecules. Instead, it is likely that thewashing steps remove the majority of particlesthat are only lightly attached to the carbon film,leaving behind only particles that have a substantialinteraction surface with the carbon layer. Imagestaken from specimens stained with ammoniummolybdate showed molecules in random anduniformly distributed orientations (Figure 5(c)). Inthis case the sample was pre-mixed with theglucose-containing staining solution, and the gridwas not washed after application of the sample.

The orientations in which a molecule will adsorbto a grid will always depend on the chemical andmorphological characteristics of the molecule itself,but based on our experiments we can identifyfactors that influence the adsorption behavior ofmolecules. Our observations suggest that formolecules adsorbed to a carbon film, washingsteps and complete drying of the grid are factorsthat tend to favor preferred orientations, whereaspre-mixing the sample with a stain/sugar solutionand freezing of the grid are factors that tend to favorrandom orientations.

Flattening and distortions introduced by thespecimen preparation technique

The density map produced by random conical tiltusing side view particles of Tf–TfR complex preparedby the conventional negative staining techniquesuffers from substantial preparation-induced arti-facts. One side of the unsymmetrized reconstructionis flat (Figure 2(d), left side in views 2 to 4). Also, inone of the Tf molecules the two lobes are closertogether, and in this Tf molecule the Tf C-lobe isnotably smaller than in the other Tf molecule (markedby an asterisk in views 1 and 4 of Figure 2(d)). Inaddition, the density representing the TfR dimershows no indication of its characteristic butterflyshape. Finally, the reconstruction is severely flattened.

Such deformations induced by negative staining arewell known and have already been analyzed by anumber of investigators.31,32

A flat surface on one side of the reconstruction isseen in the reconstruction from negatively stainedparticles and, to a lesser degree, in the unsymme-trized reconstructions from both the carbon sand-wiched (Figure 3(d)) and the cryo-negativelystained particles (Figure 4(d)). The flat surfacemost likely represents the interface of the complexwith the carbon layer. This distortion can probablynot be avoided, but it is minimal in samplesprepared by cryo-negative staining. The two Tfsare bound to the TfR dimer off-axis, so that whenthe complex adsorbs to the carbon film, one of theTfs will come into contact with the carbon film moreextensively than the other one. This is most likelythe reason for the observed asymmetry between thetwo Tf molecules, as the Tf making more contactwith the carbon film is likely to be more deformedthan the other Tf. Indeed, the Tf molecule with themore closely spaced lobes and the small C-lobe isclose to the flattened surface of the reconstruction(Figure 2(d), view 4). Similar distortions can also beseen in the unsymmetrized reconstructions fromcarbon sandwiched (Figure 3(d)) and cryo-nega-tively stained particles (Figure 4(d)), but the effect ismore subtle and the C-lobe of the Tf molecule closeto the carbon film does not appear significantlysmaller than that of the other Tf.

The density representing the TfR shows thestrongest deformations when prepared by theconventional negative staining protocol and com-pletely lacks any indication of a bipartite organi-zation or the characteristic butterfly shape of thereceptor dimer (Figure 2(d), view 2). There areseveral possible explanations for the lack ofstructural detail in the TfR density, such as themissing cone problem, structural collapse of thereceptor upon drying of the grid, and incompleteembedding of the receptor in the stain layer. Themissing cone problem results from the fact thatspecimens can only be tilted to a limited angle in theelectron microscope, typically not more than 608,which prevents the sampling of a cone-shapedvolume in Fourier space. This leads to an aniso-tropic resolution of the reconstructed density mapwith the resolution in the direction perpendicular tothe carbon film being lower than the in-planeresolution. Both the unsymmetrized maps fromcarbon sandwiched (Figure 3(d)) and cryo-nega-tively stained particles (Figure 4(d)) were recon-structed from data collected in the same way as thedata for the particles negatively stained by theconventional method. These density maps showclear indications for both TfRs in the dimer. Wetherefore exclude the missing cone as the cause forthe loss of structural detail in the density represent-ing the TfR dimer. The lack of structuraldetail probably results from the combined effectof incomplete stain embedding and structuralcollapse. In particles adsorbed to the carbonpresenting the side view, the TfR constitutes the


density furthest removed from the carbon support.It is therefore conceivable that the TfR in the dimerfurther removed from the carbon film protrudesfrom the stain layer and is thus “invisible”. Inaddition, the apical domains, which represent thetips of the butterfly wings, are only loosely attachedto the remainder of the TfR molecules and thereforemay easily be displaced.22 It is quite likely that theapical domain of the TfR molecule further awayfrom the carbon film collapses on top of the apicaldomain of the TfR in contact with the carbon film.Although the unsymmetrized density maps calcu-lated from carbon sandwiched (Figure 3(d)) andcryo-negatively stained particles (Figure 4(d)) showmore of the butterfly shape of the receptor dimer,they also reveal distortions. In the reconstructionfrom the carbon sandwiched molecules, thereceptor in contact with the carbon film (left sidein view 2 of Figure 3(d)) seems to have slid downrelative to the TfR not in contact with the carbon.The distortions seen in the density maps calculatedfrom the carbon sandwiched and cryo-negativelystained particles thus may again represent effects ofthe adsorption of the molecules to a carbon film.

The use of a second carbon film covering thesample clearly has a beneficial effect on the outcomeof the 3D reconstruction (compare Figure 3(d) withFigure 2(d)). This is most likely caused by a betterembedding of the particles in the stain layer, whichis the reason why the carbon sandwich techniquewas originally introduced.23 The complete stainembedding is reflected by the lack of a stain shadowbelow the particles in images of tilted specimens,which is advantageous, because the strong contrastbetween the stain and the particle may influence thealignment of the particle images. Since this strongfeature depends solely on the orientation of the tiltaxis rather than on the inherent structure of themolecules, the stain cloud could interfere with thealignment of the particles according to their lessstrongly contrasted inherent features.

The distance between the two carbon filmssandwiching the particles is crucial. Too big adistance results in a thick stain layer, in which theparticles are difficult to see, whereas too small adistance leads to squashing of the particles. Thissquashing is different from the simple flattening ofmolecules upon drying of the specimen, since theformer leads to projection averages that show theparticles to be enlarged, whereas the latter has onlylittle influence on the appearance of the particles inprojection averages: 3D reconstructions of squashedparticles are shown in Figures 3(g) and 4(g), whichshow a spread-out appearance of the molecule. Thedensity map of the squashed molecules from theimages of carbon sandwiched preparations(Figure 3(g), view 2) is also slightly more flattenedthan that of the unsquashed particle (Figure 3(d),view 2). This is consistent with the two approachingcarbon films exerting force on the trappedmolecules. The density map of the squashedmolecules obtained with the cryo-negativelystained particles does not, however, show such an

increased flattening (Figure 4(g), view 2) comparedto the unsquashed particle (Figure 4(d), view 2).This comparison indicates that the squashing of theparticles by the two carbon layers does not simplyreflect mechanical force experienced by the sand-wiched particles.

Symmetry as low as 2-fold in the imagedmolecule can substantially improve the densitymap. Many distortions seen in the unsymmetrizedmaps of the side view particles (Figures 2(d), 3(d),and 4(d)) were corrected, at least to some degree, byapplying the 2-fold symmetry of the Tf–TfRcomplex to the density map (Figures 2(e), 3(e),and 4(e)). 2-fold symmetrization does not, however,correct for the substantial flattening seen in thereconstructions from the top view particles (Figures2(f), 3(f), and 4(f)), which do not look significantlydifferent from the unsymmetrized reconstructions(data not shown). The reconstructions from the topview particles thus make the point that applicationof symmetry to single particle reconstructions isonly helpful in reducing distortion artifacts if themolecule does not adsorb to the carbon support filmwith its symmetry axis oriented perpendicular tothe carbon film.

Staining with an ammonium molybdate/glucosemixture produced images with low contrast(Figure 5(a)), and the particles were adsorbed tothe carbon film in random orientations (Figure 5(c)).Because of the poor SNR of the images, the classaverages (Figure 5(b)) also appeared noisier thanthose obtained from images of vitrified samples(Figure 1(b)). This might be the reason why wefailed with this data set to produce a reasonableinitial 3D map with the angular reconstitutionapproach and had to align the class averages tothe density map obtained with the images of thevitrified specimen. The final 3D map shows fewdistortions (Figure 5(d)). Absence of distortionsdoes not mean, however, that the particles areunflattened. As seen, for example, in the projectionaverages from the conventionally negativelystained particles (Figure 2(c)), flattening due tospecimen drying does not show up in projections.Since we had uniformly distributed views of theparticles, and therefore could not calculate arandom conical tilt 3D reconstruction, we cannotjudge the extent of flattening introduced by thisspecimen preparation technique. Since the imagesshowed less contrast than those of uranyl formatestained specimens (ideal for random conical tilt) oreven vitrified specimens (ideal for angular recon-stitution), we do not view staining with a glucose-containing ammonium molybdate solution as anadvantageous specimen preparation technique.

Cryo-negative staining appears to produce the best3D reconstructions displaying the least distortionsand flattening artifacts, especially if symmetry can beapplied to the density map as in the case of the Tf–TfRcomplex. Preparation and imaging of cryo-negativelystained samples is not straightforward, however.Often grids prepared in this way show the particles tobe positively stained, making these grids useless for


data collection. In addition, different areas of the gridcan vary greatly, showing areas with stain layers toothin or too thick. Even in areas that appear to have astain layer of optimal thickness, many particles aresquashed and are thus unsuitable for 3D reconstruc-tion. Moreover, cryo-negatively stained samples areparticularly radiation sensitive, so that many imagesshow beam-induced bubbles. All these effectscombined account for a very low yield of goodparticles that can be used for 3D reconstruction. Wethink that the improved quality of the resultingdensity maps compensates for the increased amountof work needed to assemble a large enough data set.We note, however, that even with the cryo-negativestaining procedure, reliable density maps are onlyobtained in favorable cases. Under unfavorableconditions, particles prepared by this technique canproduce badly distorted and flattened reconstruc-tions as illustrated by the density map obtained withthe top view particles (Figure 4(f)).

Influence of the reference model on the finalreconstruction

We believe that the most reliable initial 3Ddensity map would be obtained by calculating arandom conical tilt reconstruction using tilt pairstaken from vitrified specimens. This approachcombines the advantage of specimen vitrification,i.e. few preparation artifacts, with the reliability ofthe random conical tilt 3D reconstruction technique.In some favorable cases this approach is feasible,such as with molecules that are large enough andadopt preferred orientations in the vitreous icelayer, as in the case of the ryanodine receptor.32

Collecting tilt pairs of vitrified specimens isextremely difficult, however, and for the ryanodinereceptor only nine out of several hundred imagepairs were of sufficiently high quality to be usefulfor image processing.32 Despite this technicalchallenge we attempted to record tilt pairs of thevitrified Tf–TfR complex, but the small size of thecomplex made it almost impossible to see themolecules in images taken from tilted specimenand to correlate them to the corresponding particlesin the images of the untilted specimen.

We therefore tested whether random conical tiltreconstructions obtained with specimens pre-pared by various negative staining techniquescould be used as models to guide the 3Dreconstruction using images taken from vitrifiedspecimens. We chose to align class averagesrather than raw images to the reference models,because of their better SNRs. Furthermore, tomake the results independent of CTF effects, welow pass-filtered the raw images to a resolutionof 40 A before calculating the class averages thatwere used for the alignment to the referencemodels. We did not refine the resulting densitymaps to high resolution, because our maininterest here was the overall shape of the densitymaps and, in particular, how the shape of the

density map was influenced by that of thereference model.

We first used our best random conical tilt 3Dreconstruction, the 2-fold symmetrized density mapobtained with the cryo-negatively stained side viewparticles (Figure 4(e)), and FREALIGN produced adensity map with a very accurate shape(Figure 6(b)). To simulate an asymmetric molecule,we used the unsymmetrized density map obtainedwith the cryo-negatively stained side view particles(Figure 4(d)) as reference model and abstained fromusing the 2-fold symmetry in the refinementprocess. The resulting density map did not show aperfect 2-fold symmetry (Figure 6(c)), but thedensity map was clearly improved over thereference model. This result strongly suggests thatthe refinement procedure can, at least to somedegree, overcome distortions present in the refer-ence model. This was not the case, however, whenwe used the 2-fold symmetrized density mapobtained with the cryo-negatively stained topview particles as reference model, even when2-fold symmetry was enforced during the refine-ment (Figure 4(f)). In this case, the resulting densitymap looked only very remotely similar to thestructure of the Tf–TfR complex (Figure 6(d)). Theunrecognizable density map is most likely due tomisalignment of many class averages to the almostfeatureless reference map. When we used the 2-foldsymmetrized density map obtained with theconventionally negatively stained top view par-ticles as reference model, we produced a similarlyunrecognizable reconstruction (data not shown).A density map resulting from refinement withFREALIGN that looks very different from the initialreference map may thus indicate that the initialmodel was inaccurate, e.g. because it suffered fromsubstantial distortions and flattening artifacts.Further studies are then needed to determinewhether the “refined” density map is indeed agood representation of the true structure of themolecule. Finally, we used the unsymmetrized and2-fold symmetrized density maps obtained with theconventionally negatively stained side view par-ticles (Figure 2(d) and (e)) and performed therefinement without and with enforced 2-foldsymmetry, respectively. Both refinements produceda density map (Figure 6(e) and (f)) that wassignificantly improved over the respective referencemodel in terms of flattening and asymmetry. Theseresults indicate that even a significantly flattenedreference model can be used to align the classimages from vitrified specimens, as long as it showsa sufficient amount of correct structural detail towarrant successful matching of the averages to thereference model.

Conclusions

Our results show that 3D reconstructionsobtained by single particle EM are often far fromperfect. This finding emphasizes the need for very


careful analysis of the density maps produced bythis technique, especially if molecules withunknown structures are studied. The angularreconstitution approach, commonly used to analyzeimages of vitrified specimens, can potentiallyproduce inaccurate reconstructions, although ahigh symmetry of the molecule under investigationusually overcomes this caveat. In the case ofmolecules with low or no inherent symmetry, 3Dmaps obtained with the angular reconstitutionapproach should be viewed with great skepticismand not be trusted blindly unless confirmed byother structural information, such as crystal struc-tures or independently determined 3D reconstruc-tions. The random conical tilt approach producesunambiguous reconstructions. Since this approachrequires the molecules to adopt preferred orien-tations, which is often only seen in negativelystained specimens, the density maps are oftenaffected by distortions and flattening artifacts dueto the need to adsorb the molecules to a carbon filmand to dry the specimen. Even in the mostsophisticated negative staining procedures, suchas cryo-negative staining, distortions cannot beruled out. Although the distortions cannot bepredicted, they can be understood, as they primar-ily affect features of the molecule perpendicular tothe carbon film. Random conical tilt reconstructionsof negatively stained specimens can therefore beused to validate or discount density maps obtainedby angular reconstitution of vitrified specimens.Alternatively, a random conical tilt reconstructionobtained with a cryo-negatively stained specimencan be used as reference model to align imagestaken from vitrified specimens. The programFREALIGN is fast and well suited for this purpose.While the reference model can introduce subtleartifacts in the final reconstruction obtained withthe ice data, distortions and flattening artifacts ofthe reference model are largely removed during therefinement process. This procedure fails, however,if the reference model is substantially distorted.A large difference between the reference model andthe final map is thus an indication that the chosenreference model may not have been a goodrepresentation of the molecular structure.

Materials and Methods

Specimen preparation and electron microscopy

Recombinant human Tf–TfR complex was prepared asdescribed9 and diluted to a concentration of about0.01 mg/ml. Vitrification of Tf–TfR complex was doneas described.9 Protocols for conventional negative stain-ing, the carbon sandwich technique, ammonium molyb-date staining, and cryo-negative staining have beendescribed.13 Images of vitrified specimens were recordedwith a Tecnai F20 electron microscope as described.9

Specimens prepared by conventional negative stainingand the carbon sandwich technique were imaged at roomtemperature in a Gatan 670 ultra-high tilt holder, whereasspecimens prepared by ammonium molybdate and

cryo-negative staining were imaged at liquid nitrogentemperature of about K180 8C in an Oxford cryo-transferholder. Samples prepared with ammonium molybdatewere imaged at 08 tilt, while samples prepared byconventional negative staining and the carbon sandwichtechnique were imaged at tilt angles of 608 and 08. In the caseof cryo-negative staining the sample was imaged at tiltangles of 508 and 08. All images were taken with an FEITecnai T12 microscope equipped with an LaB6 filament andoperated at 120 kV. Images were taken at a magnification of52,000! on Kodak SO-163 film using low-dose proceduresand developed for 12 min with full-strength Kodak D-19developer at 20 8C. All micrographs were visually inspectedwith an optical laser diffractometer and only drift-freeimages were digitized with a Zeiss SCAI scanner using astep size of 7 mm. The digitized images were furtheraveraged over 3!3 pixels to give a final pixel size of4.04 A/pixel at the specimen level.

Image processing

Images of vitrified Tf–TfR complex prepared withholey carbon film were processed as described.9 Briefly,a total of 36,266 particles were selected interactively from196 micrographs using Ximdisp, the display programassociated with the MRC program suite.33 Using theIMAGIC software package16 the windowed particleswere band-pass filtered to a resolution of 170 A to 40 A,followed by five cycles of MRA, MSA and classificationinto 500 classes. A total of 124 class averages were selectedout of the final 500 class averages. Euler angles wereassigned to these class averages using the AngularReconstitution command in IMAGIC with the NEW_PROJECTION/FRESH option. The class averages withthe assigned Euler angles were then used to calculate a 3Dreconstruction. Projections calculated from the 3D recon-struction at an angular interval of 158 were used as anchorset to further refine the Euler angles of the class averages.After two cycles of Euler angle refinement using anchorsets, 157 additional class averages were added into thedata set. A total of six cycles of Euler angle refinementwas performed and the final 3D reconstruction shown inFigure 1(d) calculated from a total of 281 class averages.

Images of vitrified Tf–TfR complex adsorbed tocontinuous carbon film were recorded and digitized inthe same way as for the vitrified Tf–TfR complex on holeycarbon film. A total of 21,719 particles were selectedinteractively from 46 micrographs. The particles wereband-pass filtered as before, followed by only one cycle ofMRA, MSA and classification into 500 classes. The refinedmodel of the Tf–TfR complex9 was filtered to 10 A andused to calculate projections at an angular interval of 158.These projections were used as anchor set to assign Eulerangles to all 500 class averages. The resulting 3D map wasused as reference model to which the raw particle imageswere aligned with FREALIGN. The density map wasrefined to 7.5 A.

For images of the Tf–TfR complex stained withammonium molybdate, a total of 17,107 particles wereinteractively selected from 39 micrographs. The particleswere band-pass filtered and subjected to five cycles ofMRA, MSA and classification. As for the vitrified Tf–TfRcomplex on continuous carbon film, the Euler angles ofthe final 500 class averages were determined using theanchor set created from the refined 3D model of thevitrified Tf–TfR complex on holey carbon film, andfurther refined with FREALIGN.

All tilt pairs were processed using the SPIDER softwarepackage.17 Particles were selected interactively using


WEB, the display program associated with SPIDER. Theparticles from the images of the untilted specimens weresubjected to eight cycles of multi-reference alignment andclassification into 50 classes. Class averages showing thesame projection structure were combined for furtherprocessing. The initial 3D reconstructions were calculatedwith the particles selected from the images of the tiltedspecimens using the random conical approach:10 10% ofthe particles from the corresponding classes from theimages of the untilted specimens were then added to thedata set and the structures refined by ten cycles of angularrefinement, which is based on projection matching.The final resolutions of the 3D reconstructions wereestimated from the Fourier shell correlation (FSC) curveusing the FSCZ0.5 cut-off criterion.

For the Tf–TfR complex prepared by the conventionalnegative staining procedure, a total of 8410 pairs ofparticles were selected from 84 pairs of micrographs. Thereconstructions shown in Figure 2 contain 1461 particles(Figure 2(d) and (e)) and 1412 particles (Figure 2(f)). Theresolutions for the 3D reconstructions shown inFigure 2(e) and (f) were determined as 27 A.

For the Tf–TfR complex prepared by the carbon sandwichtechnique, a total of 14,820 pairs of particles were selectedfrom 54 pairs of micrographs. The reconstructions shown inFigure 3 contain 5163 particles (Figure 3(d) and (e0), 468particles (Figure 3(f)), and 512 particles (Figure 3(g)). Theresolutions for the 3D reconstructions shown in Figure 3(e),(f), and (g) were determined as 22 A, 30 A, and 27 A.

For the Tf–TfR complex prepared by cryo-negativestaining, a total of 23,142 pairs of particles were selectedfrom 48 pairs of micrographs. The reconstructions shownin Figure 4 contain 4190 particles (Figure 4(d) and (e)), 380particles (Figure 4(f)), and 424 particles (Figure 4(g)). Theresolutions for the 3D reconstructions shown inFigure 4(e), (f), and (g) were determined as 24 A, 29 A,and 29 A.

FREALIGN was used to assign Euler angles to the 500class averages of the vitrified Tf–TfR complex. The 3Dreconstructions obtained with the random conical tiltapproach were used as initial reference models. Option 3in FREALIGN (simple search and refine) was used with a158 angular step size and a refinement limitation of 20 A.

Docking of the atomic structure of the Tf–TfR complex(pdb code: 1SUV) into the density maps was donemanually using the program Chimera.34

Acknowledgements

We thank Pawel Penczek for many valuablesuggestions. This work was supported by NationalInstitutes of Health Grant GM62580 (to DavidDeRosier). The molecular EM facility at HarvardMedical School was established by a generousdonation from the Giovanni Armenise HarvardCenter for Structural Biology.

References

1. Bottcher, B., Wynne, S. A. & Crowther, R. A. (1997).Determination of the fold of the core protein ofhepatitis B virus by electron cryomicroscopy. Nature,386, 88–91.

2. Zhou, Z. H., Dougherty, M., Jakana, J., He, J.,Rixon, F. J. & Chiu, W. (2000). Seeing the herpesviruscapsid at 8.5 A. Science, 288, 877–880.

3. Gabashvili, I. S., Agrawal, R. K., Spahn, C. M.,Grassucci, R. A., Svergun, D. I., Frank, J. & Penczek, P.(2000). Solution structure of the E. coli 70S ribosome at11.5 A resolution. Cell, 100, 537–549.

4. Stark, H., Rodnina, M. V., Wieden, H. J., van Heel, M.& Wintermeyer, W. (2000). Large-scale movement ofelongation factor G and extensive conformationalchange of the ribosome during translocation. Cell, 100,301–309.

5. Jurica, M. S., Sousa, D., Moore, M. J. & Grigorieff, N.(2004). Three-dimensional structure of C complexspliceosomes by electron microscopy. Nature Struct.Mol. Biol. 11, 265–269.

6. Boehringer, D., Makarov, E. M., Sander, B.,Makarova, O. V., Kastner, B., Luhrmann, R. &Stark, H. (2004). Three-dimensional structure of apre-catalytic human spliceosomal complex B. NatureStruct. Mol. Biol. 11, 463–468.

7. Stark, H., Dube, P., Luhrmann, R. & Kastner, B. (2001).Arrangement of RNA and proteins in the spliceo-somal U1 small nuclear ribonucleoprotein particle.Nature, 409, 539–542.

8. Orlova, E. V., Rahman, M. A., Gowen, B., Volynski, K. E.,Ashton, A. C., Manser, C. et al. (2000). Structure of alpha-latrotoxin oligomers reveals that divalent cation-dependent tetramers form membrane pores. NatureStruct. Biol. 7, 48–53.

9. Cheng, Y., Zak, O., Aisen, P., Harrison, S. C. & Walz, T.(2004). Structure of the human transferrin receptor–transferrin complex. Cell, 116, 565–576.

10. Radermacher, M., Wagenknecht, T., Verschoor, A. &Frank, J. (1987). Three-dimensional reconstructionfrom a single-exposure, random conical tilt seriesapplied to the 50 S ribosomal subunit of Escherichiacoli. J. Microsc. 146, 113–136.

11. van Heel, M. (1987). Angular reconstitution: a poster-iori assignment of projection directions for 3Dreconstruction. Ultramicroscopy, 21, 111–123.

12. Bremer, A., Henn, C., Engel, A., Baumeister, W. &Aebi, U. (1992). Has negative staining still a place inbiomacromolecular electron microscopy? Ultramicro-scopy, 46, 85–111.

13. Ohi, M., Li, Y., Cheng, Y. & Walz, T. (2004). Negativestaining and image classification–powerful tools inmodern electron microscopy. Biol. Proc. Online, 6,23–34.

14. Adrian, M., Dubochet, J., Lepault, J. & McDowall,A. W. (1984). Cryo-electron microscopy of viruses.Nature, 308, 32–36.

15. van Heel, M., Winkler, H. P., Orlova, E. V. & Schatz, M.(1992). Structural analysis of ice-embedded singleparticles. Scanning Microsc. Suppl. 6, 23–42.

16. van Heel, M., Harauz, G., Orlova, E. V., Schmidt, R. &Schatz, M. (1996). A new generation of the IMAGICimage processing system. J. Struct. Biol. 116, 17–24.

17. Frank, J., Rademacher, M., Penczek, P., Zhu, J.,Ladjadj, M. & Leith, A. (1996). SPIDER and WEB:processing and visulization of images in 3D electronmicroscopy and related fields. J. Struct. Biol. 116,190–199.

18. Ludtke, S. J., Baldwin, P. R. & Chiu, W. (1999). EMAN:semiautomated software for high-resolution single-particle reconstructions. J. Struct. Biol. 128, 82–97.

19. Stewart, A. & Grigorieff, N. (2004). Noise bias in therefinement of structures derived from single particles.Ultramicroscopy, 102, 67–84.


20. Penczek, P. A., Zhu, J. & Frank, J. (1996). A common-lines based method for determining orientations forNO3 particle projections simultaneously. Ultramicro-scopy, 63, 205–218.

21. Hall, D. R., Hadden, J. M., Leonard, G. A., Bailey, S.,Neu, M., Winn, M. & Lindley, P. F. (2002). The crystaland molecular structures of diferric porcine andrabbit serum transferrins at resolutions of 2.15 and2.60 A, respectively. Acta Crystallog. sect. D, 58, 70–80.

22. Lawrence, C. M., Ray, S., Babyonyshev, M., Galluser,R., Borhani, D. W. & Harrison, S. C. (1999). Crystalstructure of the ectodomain of human transferrinreceptor. Science, 286, 779–782.

23. Stoffler-Meilicke, M. & Stoffler, G. (1988). Localizationof ribosomal proteins on the surface of ribosomalsubunits from Escherichia coli using immunoelectronmicroscopy. Methods Enzymol. 164, 503–520.

24. Golas, M. M., Sander, B., Will, C. L., Luhrmann, R. &Stark, H. (2003). Molecular architecture of the multi-protein splicing factor SF3b. Science, 300, 980–984.

25. Adrian, M., Dubochet, J., Fuller, S. D. & Harris, J. R.(1998). Cryo-negative staining. Micron, 29, 145–160.

26. Orlova, E. V., Dube, P., Harris, J. R., Beckman, E.,Zemlin, F., Markl, J. & van Heel, M. (1997). Structureof keyhole limpet hemocyanin type 1 (KLH1) at 15 Aresolution by electron cryomicroscopy and angularreconstitution. J. Mol. Biol. 271, 417–437.

27. Unwin, P. N. & Henderson, R. (1975). Molecularstructure determination by electron microscopy of

unstained crystalline specimens. J. Mol. Biol. 94,425–440.

28. Gonen, T., Sliz, P., Kistler, J., Cheng, Y. & Walz, T.(2004). Aquaporin-0 membrane junctions reveal thestructure of a closed water pore. Nature, 429, 193–197.

29. Rosenthal, P. B. & Henderson, R. (2003). Optimaldetermination of particle orientation, absolute hand,and contrast loss in single-particle electron cryomi-croscopy. J. Mol. Biol. 333, 721–745.

30. Grigorieff, N. (1998). Three-dimensional structure ofbovine NADH:ubiquinone oxidoreductase (complex I)at 22 A in ice. J. Mol. Biol. 277, 1033–1046.

31. Boisset, N., Taveau, J. C., Lamy, J., Wagenknecht, T.,Radermacher, M. & Frank, J. (1990). Three-dimen-sional reconstruction of native Androctonus australishemocyanin. J. Mol. Biol. 216, 743–760.

32. Radermacher, M., Rao, V., Grassucci, R., Frank, J.,Timerman, A. P., Fleischer, S. & Wagenknecht, T.(1994). Cryo-electron microscopy and three-dimen-sional reconstruction of the calcium release channel/ryanodine receptor from skeletal muscle. J. Cell. Biol.127, 411–423.

33. Crowther, R. A., Henderson, R. & Smith, J. M. (1996).MRC image processing programs. J. Struct. Biol. 116,9–16.

34. Pettersen, E. F., Goddard, T. D., Huang, C. C., Couch,G. S., Greenblatt, D. M., Meng, E. C. & Ferrin, T. E.(2004). UCSF Chimera–a visualization system forexploratory research and analysis. J. Comput. Chem.25, 1605–1612.

Edited by W. Baumeister

(Received 3 October 2005; received in revised form 31 October 2005; accepted 7 November 2005)Available online 28 November 2005

Documents

Single Particle Reconstructions of the Transferrin– Transferrin