ELSEVIER SCIENCE IRELAND

Chemistry and Physics of Lipids 67/68 (1994) 81-89

Chemistry and Physics of

LIPIDS

Electron cryomicroscopy and digital image processing of lipoprotein(a)

Jacqueline Sines *a'b, Rosalba Rothnagel c, Marin van Heel d, John W. Gaubatz b, • a b c • ac Joel D. Mornsett " , Wah Chin '

aW.M. Keck Center for Computational Biology Houston. TX 77030. USA bThe Department of Medich~e, Baylor College of Medicine, 6550 Fannin, MS A601, Houston, TX 77030, USA

"Verna and Marts McLean Department of Biochemistry, Baylor College of Medicine. One Baylor Plaza, Houston, TX 77030, USA dFritz Haber Institute of the Max Planck Society, Faradayweg 4-6. D-IO00 Berlin 33, German)'

(Accepted 12 November 1992)

Abstract

Electron cryomicroscopy was used to study the structure of human lipoprotein(a) (Lp(a)), a plasma lipoprotein implicated in cardiovascular disease. An individual Lp(a) particle consists of a neutral lipid core within a shell of phos- pholipid, cholesterol and glycoprotein. In principle, electron cryomicroscopy images of single particles should contain structural detail attributable to the density differences among these components and the surrounding buffer. We observed such structural detail in images of frozen, hydrated Lp(a) particles. Lp(a) particles appeared to be roughly spherical in shape with an average diameter of 210 ~,. As is generally true for unstained samples in vitreous ice, imaged with a low electron dose, these images have low contrast with low signal-to-noise ratios. To increase the signal-to-noise ratio, we averaged classes of similar particles. We began with a set of 5813 randomly oriented Lp(a) particles and generated classes using a linear multivariate statistical method, followed by hierarchical ascendent classification. Our initial classification, based on only the first eight eigenvectors, separated particles on the basis of gross size and shape. After a rough reference-free alignment step, a second classification used the finer details in the images. This approach yielded class averages with structural detail only faintly visible in the raw, single images.

Key words: Lipoprotein(a); Lp(a); Cryomicroscopy; Electron microscopy; Image processing; Multivariate statistical an- alysis; MSA; Single particles

* Corresponding author, Verna and Marrs McLean Depart- ment of Biochemistry, Baylor College of Medicine, One Baylor Plaza, Houston, TX 77030, USA. Abbreviations: Lp(a); lipoprotein (a); apo(a), apolipoprotein (a); apo B-100, apolipoprotein B-100; LDL, low density lipo- protein.

I. Introduction

Lipoprote in(a) (Lp(a)) is a p lasma l ipoprote in which has been impl icated in co ronary ar tery disease. The basic structure o f the par t ic le is a core of apola r l ipid enclosed in a p redominan t ly phos-

0009-3084/94/$07.00 © 1994 Elsevier Science Ireland Ltd. All rights reserved. SSD1 0009-3084(93)02227-1

82 J. Sines et al./ Chem. Phys. Lipids 67/68 (1994) 81-89

pholipid monolayer with associated glycoproteins. (For recent review, see Morrisett et al. (1987, 1990), Loscalzo (1990), Scanu and Fless (1990), Lawn (1991), and Scanu (1992).) The molecular weight of a particle is approximately 5000 kDa. About half of this mass is core lipid, mostly choles- teryl ester with smaller amounts of triglyceride and cholesterol. Monolayer components, phospholipid and cholesterol in a 2:1-3:1-ratio, make up about 30% of the particle by weight. The remaining parti- cle mass is attributable to two glycoproteins, apolipoprotein B-100 (apo B-100) and apolipopro- tein(a) (apo(a)), each usually present at one copy per particle, although particles with two copies of apo(a) have been reported (Scanu, 1992.) Various techniques have yielded estimated particle sizes ranging in diameter from 210 to 262 A, (Morrisett et al., 1987).

Apo(a) and apo B-100 appear to be covalently linked by at least one disulfide bond. After reduc- tion of this bond, apo(a) may be removed from the particle by ultracentrifugation or electrophoresis but not by gel filtration. This leaves a particle, Lp(a-), which is very similar to low density lipo- protein (LDL). Apo B-100 requires more vigorous treatment, such as strong denaturants, to dissoci- ate it from the Lp(a) particle. Both apoproteins are large; apo B-100 is about 500 kDa in molecular weight, while apo(a) varies from about 418 to 838 kDa. Glycosylation levels are 5-10% of the mass in the case of apo B-100 and 30-40% for apo(a). Many of the apo(a) sugar residues are sialic acid, generating a high negative charge for both the iso- lated glycoprotein and the Lp(a) particle. Apo(a) has been sequenced and found to exhibit extensive homology to plasminogen (McLean, 1987). It is composed of an N-terminal kringle domain and a C-terminal serine protease domain. Eleven dif- ferent kringle types are present, with all except the second present in single copies. Lp(a) kringle 2 is found in a variable number of copies, accounting for most of the molecular weight variability of apo(a). Between 4 and 28 copies have been reported, with 28 in the polymorph which was se- quenced (Scanu and Fless, 1990). A single donor can have one or two polymorphs of apo(a) as determined from the number of bands on elec- trophoresis.

Various components of Lp(a) differ substantial- ly in density, with the lipid core of the particle being of lower density than the surrounding buffer and the phospholipid head groups and glycopro- teins in the outer shell being higher. It should be possible to exploit these density differences to pro- vide contrast in images obtained by electron cryomicroscopy. In this technique, rapid, very cold freezing of the sample produces vitreous ice, a non-crystalline solid which is believed to retain the native hydrated state of the specimen. (See Chiu, (1986) and Dubochet et al. (1988) for review.) In the absence of stain, all contrast is produced by the sample, with the image being a two-dimensional projection of the three-dimensional density of the particle in buffer. We present here the first electron microscopy images of Lp(a). These images were produced using unstained sample in vitreous ice. We also present the results of computer processing of 5813 single Lp(a) particle images. The analysis used a combination of a linear multivariate statistical analysis (MSA) (van Heel and Frank, 1981; Greenacre, 1984; Lebart et al., 1984), hierar- chical ascendent classification (van Heel, 1989; Frank, 1990) and alignment techniques. The resul- tant averaged particle images show reproducible features which are more distinct than the faint fea- tures of the original images.

2. Methods

2.1. Preparation and cryomicroscopy of Lp(a) Lp(a) purification generally followed the proce-

dure outlined by Gaubatz et al. (1986). Sequential ultracentrifugation of plasma from a single band phenotype donor (band a3 reported in Gaubatz et al. (1990)) yielded the fraction with density 1.06- 1.12 g/ml. After dialysis against 15 mM NaCI, 1 mM EDTA, 0.01% NAN3, 50 mM phosphate (pH 7.5), this fraction was load onto a lysine- Sepharose column, washed with 50 mM phosphate (pH 7.5) and eluted with 100 mM e-amino-n- caproic acid, 500 mM NaC1, 50 mM phosphate (pH 7.5). After dilution into ammonium acetate buffer, final concentrations were 100/~g/ml Lp(a) protein, 94 mM ammonium acetate, 2 mM am- monium carbonate, 0.2 mM EDTA, 25 mM e- amino-n-caproic acid, 125 mM NaCI, 12.5 mM

J. Sines et al./ Chem. Phys. Lipids 67/68 (1994) 81-89 83

phosphate (pH 7.4). We froze the Lp(a) particles by suspending this solution over a thin perforated nitrocellulose film on a copper grid, then quickly blotting and immersing the grid in liquid ethane (Dubochet et al., 1988). This rapid freezing pro- duces a thin layer of vitreous ice with suspended particles. Once frozen, the grid remained under liquid nitrogen until transfer to a JEOL-1200 elec- tron microscope. Images were captured on Kodak SO 163 film at a linear magnification of 30 000 with accelerating voltage of 100 kV under low-dose conditions ( - 5 e- /A 2).

2.2. Computer processing We selected three micrographs with similar

defocus for digitizing and computer processing. Images were scanned on a Perkin-Elmer PDS 1010 M microdensitometer with a 12.5-/~m aperture and ste.p size, yielding pixels corresponding to 4 ~, × 4 A areas of the sample grid. All computer pro- cessing was carried out on a Silicon Graphics Power Series II or a Silicon Graphics Iris Indigo. IMAGIC image processing software for electron microscopy (Image Science, 1991) accomplished most of the computation on images with the excep- tion of rotational alignment. A new alignment method was used (Sines, unpublished data) for reference-free rotational alignment of images.

2.3. Classification The data set of images was treated as a linear

system with rows corresponding to images, and columns to pixel values at specific locations within the images. We used a linear multivariate statis- tical analysis (MSA) to reduce this system to a lower dimension using a least-squares projection (van Heel and Frank, 1981; Greenacre, 1984; Lebart et al., 1984). Pixels and images were weighted by functions of the variation over active pixei locations within images and among images to increase the influence of data which showed less variation. In particular, the weights for the images and pixels are based on 'modulation distances' as defined by Borland and van Heel (1990). After MSA, a hierarchical ascendent classification (van Heel, 1989; Frank, 1990) divided the data set into classes based on distances between the images in the subspace chosen, allowing differential weight-

ing of the dimensions. The algorithm uses an iter- ative approach to find a solution with maximum interclass variance (van Heel, 1989).

A two-stage classification with intervening rotational alignment grouped similar particles together. (See van Heel et al. (1993) and references therein for other applications of simple and two- stage classifications.) The first classification of Lp(a) particles focused on gross size and shape, with little attention to fine detail. It used the coor- dinates in the eight most significant dimensions (defined by the eigenvectors with the highest eigen- values) to assign each of the 5813 particles to one of thirty classes. We averaged all images in each of the classes to produce a class average which show- ed a characteristic shape for the class. The next step was to sort the thirty classes into five groups based on these shapes of class averages. For each group, we used rotation angles determined by a rough reference-free alignment of the class averages to align the individual particle images within classes. The purpose of these steps was to produce five groups of about 1200 particles with similarity in shape and orientation. Each group was then subjected to an intragroup MSA and classification into 20 classes using the coordinates relative to the 24 most significant eigenvectors. Since each group was defined by similarity in gross features, differences between particles within each group tended to be due to finer features than in the original complete data set. This second classifica- tion accentuated these features, resulting in class averages which showed more detail than the origi- nal images.

3. Results

Fig. 1. shows frozen, hydrated Lp(a) particles collected from a donor with a single band phenotype a3 (Gaubatz et al., 1990). This is a small area from one of the three micrographs of Lp(a) particles which were used for this analysis. Because no stain was used during preparation, the contrast within the image resulted only from densi- ty differences in the sample. Mathematically, the micrograph is a projection of the three-dimen- sional structure within the sample onto a two-

84 J. Sines et al. / Chem. Phys. Lipide" 67/68 (1994) 81-89

[.-

0 :::k

o Hi

>. _

, . . .q

o ~ , .q

8 ~ 8 d

o " -

g ~

2 ~

~ E ~ 8

J. Sines et al . / Chem. Phys. Lipids 67/68 (1994) 81-89 85

dimensional plane I. In these images the grey scale is defined so that lighter areas correspond to more electron-dense components such as glycoprotein and phospholipid head groups, in contrast to darker buffer or lipid. Individual Lp(a) particles are visible as roughly circular light areas against a darker background. There is variation in shape and size from particle to particle, but the average diameter of the images is about 210 ~,. The low contrast and low signal-to-noise ratio, character- istic of unstained, frozen hydrated samples such as these, make it difficult to see any features within the particles. However, some variation in shading, indicative of components of differing density, is visible in all of the particles.

One feature noticeable in nearly all Lp(a) images is a concentration of higher density at the circum- ference of the image. This is expected from the general structure of the Lp(a) particle: an approx- imately spherical shell of high density, due to the presence of glycoprotein and phospholipid head groups, enclosing a low-density lipid core. In a two-dimensional projection, such a structure is visible as a ring, since the electron path through the edge of the particle traverses more high density volume and less low density volume than a path through the particle center. To see this structure more clearly, we averaged 5813 masked, un- filtered, particle images, then further averaged the result rotationally. This averaging produced the noise-free, light-colored ring shown in Fig. 2A. Outside this ring is darker buffer, terminated by a circular mask of intermediate density. Averaging in this way removed the noise from the image, since the noise was out of phase from image to image, while density at the circumference was rein- forced, since it was present in all particles. As we made no attempt to select images with specific fea- tures in common, any features other than this gen- eral ring shape were also out of phase from image

t One omission in the current analysis was correction for the contrast transfer function (CTF), both in treatment of Lp(a) data and in the generation of the simple shell model discussed below. The CTF defines the actual relationship between the three-dimensional structure and the two-dimensional projec- tion. See Penczek et al. (1992) for a discussion of CTF correc- tions as related to single particle reconstruction.

to image and disappeared with the noise. The image-to-image variation did have an effect of smoothing the edges of the ring, however. For comparison, we present a very simple model, the projection of a sharp-edged sphere onto a plane (Fig. 2B). The two images are very similar, as can be seen clearly in the graphs of the image profiles in Fig. 2C. The most prominent difference between the two images is the sharp interior edge of the shell in the simple model (Fig. 2B) which is smoothed out by inter-image variation in the parti- cle average in Fig. 2A. These results show that our images are consistent with the general Lp(a) model of a low density lipid core enclosed by a high den- sity layer of glycoprotein and polar head groups.

To see the faintly visible features more clearly, we next tried to select Lp(a) images which had similar features at similar locations in the images. Averaging of such a group of particles would result in reinforcement of the feature, in contrast to the situation where we averaged dissimilar par- ticles. The noise, of course, would still be reduced by the averaging as in the case with all particles discussed above. We first processed images using a Gaussian band-pass filter to accentuate the fea- tures in the resolution range of interest, about 10-200 ,~. The high-pass part of the filter also accentuated the general ring shape of the particle image. Then a two-stage classification procedure, with a rough alignment belween classifications, produced classes of similar images. Fig. 3 shows some of the resulting class averages which have in common a ring shape which is about one-third of the diameter of the particle. Taken together, the different averages which exhibit this ring suggest that it is a two-dimensional projection of a torus on the surface of the particle. Fig. 3A is a montage of five class averages which have the toroidal motif off-center with respect to the particle image. In Fig. 3B, similar features are seen in the five class averages, but in each of these cases the image of the torus is more centered in relation to the parti- cle image. The class-to-class consistency suggests that this toroidal motif is representative of some consistent structure in the Lp(a) particles them- selves. We have not aligned and averaged these images because they may not represent views of similar particles or may not be superimposable by

86

A

i) . . . . . . . . . ...... ....

J. Sines et al./Chem. Phys• Lipids 67/68 (1994) 81-89

B

0

-8

i i i

++ ++

, ~ + ~ ~ + + - ~ . + +

I I I

Fig. 2. Comparison of averaged data with spherical shell. (A) Average of 5813 Lp(a) masked particle images, further averaged rota- tionally to smooth the image. Images were normalized to a mean density of 0 within a circular mask and densities outside the mask were set to zero. (B) Projection of a spherical high density shell with low density interior onto two dimensions. Outer and inner diameters of the shell are 210 ~, and 146 ~,, respectively. Pixel values were 1.0, 1.4 and 0.95 for the surrounding medium, the shell and the interior of the shell. The images were normalized and masked as in (A) (C). Density profiles along diameters of images in (A) and (B). Connected diamonds correspond to the averaged data of Fig. 2A and unconnected crosses correspond to the model of

Fig. 2B.

a simple ro ta t ion in the plane o f the two-dimen- sional image.

4. Discussion

The Lp(a) s tructure presents a difficult p rob lem in electron microscopy and single part icle image

processing. The first quest ion is whether Lp(a) particles exhibit enough homogenei ty to be amen- able to the da ta averaging and merging which are the core of image processing o f single particles. We have begun to answer that quest ion with averaged images, such as those presented in Fig. 3, which show part icle features not easily observable in the

J. Sines et al./ Chem. Phys. Lipids 67/68 (1994) 81-89 87

III II

I I

A

B

IIII mm II i II

Fig. 3. Class averages resulting from two-stage classification with intervening alignment. Each of the ten images is an average of 50 to 74 individual particle images which classified together due to some similarity in the images. This similarity may result because the images are similar views of the same Lp(a) structure or may be fortuitous. If the averages do represent different views of the same structure, the alignment of the views relative to a frame of reference requires three angles (Eulerian angles) due to the random orienta- tion of the sample on the grid. Thus we do not try to superimpose these averages after alignment in two dimensions, as is often done when preferred orientations are assumed. (A) Off-center toroidal motifs are marked with arrows. (B) Features which are similar to

those in (A) but located closer to the center of the particle image.

raw data. Each of these images is an average of all particles within a class and shows detail which is in some way characteristic of the particles within that class. This intraclass similarity may result because the particle images within the class repre- sent similar views of similar particles. It is impor- tant to realize that a given class may also contain images of similar particles at different orientations which happen to have features in common, as well as images of dissimilar particles with features in common. The premise that these class averages represent views of an Lp(a) particle is supported by consistency in features throughout the various classes. We looked at partial sums of the classes and found that details began to appear with the in- clusion of about five particle images and that the same details were present whether we cilose par- ticles near to or far from the center of the class

(data not shown.) This intraclass homogeneity does not mean that there is strict homogeneity over the entire set of Lp(a) particles, i.e. that these class averages actually represent different views of the same Lp(a) particle structure. Our first indica- tion that there may be some homogeneity in our data is the appearance of similar features at dif- ferent orientations in different class averages, such as the toroidal motif presented in Fig. 3. However, to determine that different images really are con- sistent with a single three-dimensional structure we must apply additional statistical techniques to the field of single particle reconstruction.

Assuming Lp(a) homogeneity, or the existence of a homogeneous subset of Lp(a) particles, the image processing problem is still a difficult one. In other single particle problems, symmetry or pre- ferred orientations have usually been exploited to

88 J. Sines et aL / Chem. Phys. Lipids 67/68 (1994) 81-89

reduce the difficulty of the computational problem (Frank et al., 1981; Frank, 1989; Frank and Radermacher, 1992; van Heel et al., 1993). The alignment problem for a set of centered particles oriented in the same way relative to the electron path has only one degree of freedom, rotation around an axis parallel to the electron path. With the absence of preferred orientations in our Lp(a) samples, the problem has three degrees of rota- tional freedom, rather than one, once particles are centered. This means that all three Eulerian angles of rotation are necessary to specify an orientation of each particle relative to a frame of reference, greatly complicating the analysis. Difficulties due to this increased complexity of the problem are further exacerbated by the lack of prominent fea- tures or symmetry in Lp(a) particle images. One previous approach, classification of image in- variants (Schatz and van Heel, 1992), does not rely on preferred orientations and has been successful with images of ice-embedded specimens with pro- nounced symmetry. Although we may return to this technique as our work progresses, our initial application of invariant classification with Lp(a) particles was disappointing.

These difficulties led us to a two-tiered classification approach to analyzing our Lp(a) data. Using MSA, hierarchical ascendent classifi- cation and reference-free alignment of class averages, we grouped 5813 images of Lp(a) par- ticles into 100 classes based on their pixel-to-pixel similarity. We avoided direct alignment proced- ures with their known propensity for introducing artifact (Boekema et al., 1986). A simple one-stage classification was an option. However, in the absence of preferred orientations, Lp(a) would require a very large number of particles and far more than 100 classes to cover the conformational space. A two-stage classification procedure can greatly reduce the size of the problem. One such procedure was used to align negatively stained Bacteriophage SPP1 portal protein (Dube et al., 1993). However, the first step in the method was a multi-reference alignment, which was inappropri- ate for the low-contrast images of Lp(a). We chose instead to use two successive classifications with a reference-free alignment step between them to

reduce the classification problem to a manageable size. It should be a useful technique for other pro- jects, particularly when preferred orientations or symmetry relationships do not exist.

5. Conclusions

Frozen, hydrated Lp(a) particles appear to be roughly spherical in shape with a diameter of about 210 A and a density pattern consistent with a low-density core surrounded by a high-density shell. We observed no single preferred orientation of the Lp(a) particles on the grid and no indication of a limited number of preferred orientations. In- stead, it appears that Lp(a) particles, under these conditions, assume random orientations relative to the plane of the grid. We selected 5813 Lp(a) parti- cle images and used a two-stage classification pro- cedure to generate averaged images with enhanced contrast and increased signal-to-noise ratios. These classes are likely to contain Lp(a) particles which have assumed similar orientations, so that the class averages represent specific views of Lp(a) particles. We observed features in class averages which were visible only faintly, if at all, in single particle images and which tended to show up in multiple classes. Fig. 3 shows class averages having one such feature in common. These images appear to be differing views of a toroidal shape on the sur- face of the Lp(a) particle with a diameter of about a third of the particle diameter.

We have identified particular Lp(a) features, such as this toroidal motif, but have not yet related them to known physico-chemical characteristics of the Lp(a) particle. Labelling particles with mono- clonal antibody or gold particles, as well as com- parison of Lp(a) particles with other lipoprotein particles such as LDL, should help us to identify glycoprotein moieties. We also plan to study Lp(a) particles from different donors with differing apo(a) phenotypes to see whether the molecular weight differences are detectable in electron micrographs. Once we understand more about the structure of isolated Lp(a), we will examine Lp(a) aggregation and the binding of Lp(a) to fibrin and other macromolecules.

J. Sines et al . / Chem. Phys. Lipids 67/68 (1994) 81-89 89

6. Acknowledgements

F u n d i n g was p r o v i d e d by the W . M . K e c k F o u n -

d a t i o n , N a t i o n a l L i b r a r y o f M e d i c i n e T r a i n i n g

G r a n t No . T15 L M 0 7 0 9 3 , N I H g r a n t No .

R R 0 2 2 5 0 a n d N o . H L 3 2 9 7 1 .

7. References

Boekema, E.J., Berden, J.A., and van Heel, M.G., 1986, Struc- ture of mitochondrial FI-ATPase studied by electron microcsopy and image processing. Biochim. Biophys. Acta 851, 353-360.

Borland, L. and van Heel M., 1990. Classification of image data in conjugate representation spaces. J. Opt. Soc. Am. A 7, 601-610.

Chiu, W., 1986, Electron microscopy of frozen, hydrated bio- logical specimens. Annu. Rev. Biophys. Biophys. Chem. 15, 237-257.

Dube, P., Tavares, P. Lurz, R and van Heel, M., 1993, Bacter- iophage SPPI portal protein: a DNA pump with 13-fold symmetry. EMBO J (in presst.

Dubochet, J., Adrian, M., Chang, J.-J. et al., 1988, Cryo- electron microscopy of vitrified specimens. Q. Rev. Biophys. 21(2), 129-228.

Frank, J., 1989, Image analysis of single particles. Electron Microsc. Rev. 2, 53-74.

Frank, J., 1990, Classification of macromolecular assemblies studied as ~single" particles. Q. Rev. Biophys. 23, 281-329.

Frank, J. and Radermacher, M., 1992, Three-dimensional re- construction of single particles negatively stained or in vit- reous ice. Ultramicroscopy 46, 241-262.

Frank, J., Verschoor, A., and Boublik, M., 1981, Computer averaging of electron micrographs of 40S ribosomal subunits. Science 214, 1353-1355.

Gaubatz, J.W., Cushing, G.L. and Morrisett, J.D., 1986, Quan- titation, isolation and characterization of human lipopro- tein (a). Methods Enzymol. 129, 167-186.

Gaubatz, J.W., Ghanem, K.I., Guevara, J. Jr., Nava, M.L., Patsch, W. and Morrisett, J.D., 1990, Polymorphic forms of human apolipoprotein[a]: inheritance and relationship of

their molecular weights to plasma levels of lipoprotein[a]. J. Lipid Res., 603-613.

Greenacre, M.J., 1984, Theory and Applications of Cor- respondence Analysis. Academic Press, London.

Image Science, 1991, IMAGIC-5, distributed by Image Science Software GmbH, Berlin.

Lawn, R.M., 1991, The structure and evolution of apolipopro- tein(a), Ann. Intern. Med. 115, 209-211.

Lebart, L., Morineau, A. and Warwick, K.M. (1984) Multivariate Descriptive Statistical Analysis - - Cor- resondence Analysis and Related Techniques for Large Matrices, translated by E.M. Berry, John Wiley & Sons, New York.

Loscalzo, 1990, Lipoprotein(a) a unique risk factor for atherothrombotic disease. Arteriosclerosis 10, 672-679.

Morrisett, J.D., Guyton, J.R., Gaubatz, J.W. and Gotto, A.M., Jr., 1987, Lipoprotein(a): structure, metabolism and epidemiology, in: A.M. Gotto, Jr. (Ed.), Plasma Lipopro- teins, Elsevier Science Publishers, 129-152.

Morrisett, J.D., Gaubatz, J.W., Knapp, R.D. and Guevara, J.G., Jr., 1990, Stuctural properties of apo(a): a major apo- protein of human lipoprotein(a), in: A.M. Scanu (Ed.), Li- poprotein(a), Academic Press, 53-74.

Penczek, P., Radermacher, M. and Frank, J., 1992, Three- dimensional reconstruction of single particles embedded in ice. Ultramicroscopy 40, 33-53.

Scanu, A.M., 1992, Lipoprotein(a): its inheritance and molecu- lar basis of its atherogenetic role. Mol. Cell. Biochem. 113, 127-131.

Scanu, A.M. and Fless, G.M., 1990, Lipoprotein (a) - - hetero- geneity and biological relevance. J. Clin. Invest. 85, 1709-1715.

Schatz, M. and van Heel, M, 1992, lnvariant recognition of molecular projections in vitreous ice preparations. Ultramicroscopy 45, 15-22.

van Heel, M., 1989, Classification of very large electron micro- scopical image data sets. Optik 82, 114-126.

van Heel, M. and Frank, J., 1981, Use of multivariate statistics in analysing the images of biological macromolecules. Ultramicroscopy 6, 187-194.

van Heel, M, Winkler, H., Orlova, E. and Schatz, M., 1993, Structure analysis of ice-embedded single particles. Scan- ning Microscopy Suppl. 6 (in press).