Single-particle electron microscopy in the study of membrane protein structure

Abstract

Single-particle electron microscopy (EM) provides the great advantage that protein structure can be studied without the need to grow crystals. However, due to technical limitations, this approach played only a minor role in the study of membrane protein structure. This situation has recently changed dramatically with the introduction of direct electron detection device cameras, which allow images of unprecedented quality to be recorded, also making software algorithms, such as three-dimensional classification and structure refinement, much more powerful. The enhanced potential of single-particle EM was impressively demonstrated by delivering the first long-sought atomic model of a member of the biomedically important transient receptor potential channel family. Structures of several more membrane proteins followed in short order. This review recounts the history of single-particle EM in the study of membrane proteins, describes the technical advances that now allow this approach to generate atomic models of membrane proteins and provides a brief overview of some of the membrane protein structures that have been studied by single-particle EM to date.

Introduction

Membrane proteins constitute about one-quarter of all proteins encoded in the human genome [1], but represent about half of all drug targets [2]. Despite their great biomedical importance, as of today, membrane proteins make up only ∼7% of the structures in the Protein Data Bank. This is due to challenges related to the expression and purification of membrane proteins as well as to limitations of the techniques used for protein structure analysis, i.e. X-ray crystallography, nuclear magnetic resonance (NMR) spectroscopy and electron microscopy (EM).

For many decades, X-ray crystallography has been the preferred approach to determine the structure of soluble proteins and has yielded the vast majority of known atomic-resolution structures. With respect to membrane proteins, however, X-ray crystallography is hindered by the hydrophobic nature of the membrane-embedded domains of integral membrane proteins. The conventional way to address this issue is to solubilize membrane proteins in detergents, which then allows three-dimensional (3D) crystallization in essentially the same way as for soluble proteins. However, detergents can decrease the stability of membrane proteins, potentially leading to structural changes and even complete inactivation of the protein. Therefore, new detergents have been developed in recent years that are milder and have characteristics that are particularly well suited for 3D crystallization, such as a very low critical micellar concentration (CMC) [3]. Lipidic cubic phases (LCPs) have been developed as an alternative approach to crystallizing membrane proteins in detergent solution [4]. These 3D lipid matrices provide a more native environment for the membrane proteins than detergents and thus reduce the risk of inducing structural changes [5]. However, while soluble proteins can make crystal contacts through all surfaces, the transmembrane domains of membrane proteins in LCP crystals usually make only few crystal contacts, which are also weak, so that crystals grown in LCPs tend to be smaller and less well ordered, thus usually diffracting only to a limited resolution.

NMR spectroscopy is another well-established technique for protein structure determination at atomic resolution, which also has the unique advantage that it can be used to study dynamic processes and their changes due to ligand binding or signaling. For NMR studies, proteins have to be isotope-labeled, which is straightforward for proteins that can be expressed in E. coli or in cell-free expression systems. If proteins cannot be expressed in these systems, which is the case for many eukaryotic proteins, isotope labeling remains a challenge and new expression systems that allow isotope labeling of such proteins are still under development [6]. However, directed evolution techniques have been developed to achieve high-yield expression of G protein-coupled receptors (GPCRs) in E. coli, which allows isotope labeling for NMR studies and might be applied to other membrane proteins [7]. Commonly used NMR techniques are only suitable for small proteins, as the slow relaxation time of larger proteins broadens the resonances and increases the complexity of signal assignment. Nevertheless, solution NMR has been quite successful for β-barrel membrane proteins with up to 32 kDa [8] and homo-oligomeric helical systems of similar size [9]. Alternatively, solid-state NMR has been used to study membrane proteins reconstituted into lipid bilayers, as this technique is less affected by size limitations due to the slower relaxation time [10]. However, crowding of the many signals produced by large proteins is still an issue also for solid-state NMR techniques, making isotope labeling an absolute necessity.

Electron crystallography is an EM approach that has been developed primarily to study the structure of membrane proteins [11]. For electron crystallography, the membrane protein has to be reconstituted into a lipid bilayer, and the reconstitution conditions have to be optimized so that the protein forms a periodic array in the membrane. Since the resulting crystalline arrays typically consist of only a single layer (no stacking in the third dimension as in type 1 3D crystals of membrane proteins), these single-layered arrays are known as two-dimensional (2D) crystals. As the membrane protein in 2D crystals is surrounded by lipids, structures determined by electron crystallography should closely resemble the native structure of the protein in a biological membrane. Furthermore, if sufficiently high resolution can be achieved, structures obtained by electron crystallography of 2D crystals have the potential to reveal the surrounding lipid molecules, as illustrated, for example, by structures of aquaporin-0 (AQP0) in different lipid environments [12,13]. While electron crystallography can produce structures at resolutions comparable to those obtained by X-ray crystallography, high-resolution structures require (i) the growth of large and very well-ordered 2D crystals, and (ii) the preparation of atomically flat specimens that can be used to collect high-resolution data from highly tilted specimens, both requirements that have proven very challenging [14].

Besides the more traditional approaches to structure determination of membrane proteins, single-particle EM has recently gained substantial traction as a novel way to determine the structure of membrane proteins, even at atomic resolution [15].

Structure determination of membrane proteins by single-particle EM

Single-particle reconstruction is an EM approach that does not require crystals but can produce structures from images of individual molecules in solution. The procedure involves several steps: (i) specimen preparation by vitrification (or negative staining for low-resolution work), (ii) imaging in the electron microscope, (iii) picking and windowing the particles, (iv) correction for the contrast transfer function, (v) grouping of the particles into homogeneous 2D classes (2D classification), (vi) reconstruction of an initial 3D map, (vii) grouping of the particles into homogeneous 3D classes (3D classification), (viii) refinement of the density map(s) and finally (ix) map validation and interpretation, which at sufficiently high resolution means the building and validation of an atomic model (for a recent description of the entire procedure, see [16] and references therein).

To prepare proteins for the electron microscope, they are usually quick-frozen in liquid ethane, which creates a layer of vitrified ice that embeds the proteins and is stable in the vacuum of the electron microscope [17]. However, biological specimens are beam-sensitive, and although imaging at cryogenic temperatures reduces the effects of beam damage, when recording a single image (rather than a movie, which allows later frames to be excluded), only a limited electron dose can be used for imaging, resulting in images with a low signal-to-noise ratio. The low image contrast causes two major problems: (i) only molecules are visible in images of vitrified specimens that are large enough to create sufficient electron scattering above the scattering of the surrounding ice layer. (ii) To be able to reconstruct one (or several) 3D map(s), the projection images have to be computationally aligned to each other and grouped into classes according to the orientation (and conformation) of the imaged molecules. Low image contrast limits the alignment accuracy and thus imposes a limit on the achievable resolution. As a result, essentially only large proteins (>300 kDa) could initially be studied by single-particle cryo-EM, and the resolution of the resulting maps was usually rather low, only rarely reaching subnanometer resolution. Due to these limitations, only few large and stable membrane proteins were initially studied by single-particle cryo-EM, most notably the ryanodine receptor (RyR) and ATP synthases (see below).

Advantages of single-particle EM for the study of membrane proteins

Despite the limitations stated above, single-particle EM also has notable advantages for the study of membrane proteins, which include that (i) it only requires small amounts of proteins, (ii) it allows heterogeneous specimens to be studied, and (iii) it can easily be used to visualize larger assemblies that a membrane protein may form with other components.

Single-particle EM only requires a small amount of protein

The need for only small protein quantities is particularly advantageous for membrane proteins, which are often difficult to produce in large amounts. Preparation of a frozen grid usually requires <10 µg of purified protein, and although the number of grids needed to produce a structure depends greatly on the project, the total amount of protein needed is usually lower than the amount required for a single 2D or 3D crystallization trial (in the range of mg). However, even though the EM analysis itself does not require much protein, protein expression still needs to be sufficiently high to allow for stringent biochemical purification.

Another consideration for membrane proteins, which are solubilized by detergent, concerns the protein concentration required to perform structure analysis. Concentrating a detergent-containing protein solution usually also increases the detergent concentration, which is problematic, because high detergent concentrations can have a deleterious effect on protein stability. Crystallographic studies typically require a protein concentration of 10 mg/ml or more. In contrast, single-particle cryo-EM studies, depending on the specific membrane protein, can use protein solutions at concentrations as low as 0.3 mg/ml.

Single-particle EM can analyze the structure of a protein in different conformations

Structural heterogeneity is present in many protein samples and is often associated with functional properties of the protein. However, with the exception of NMR spectroscopy, structural techniques usually provide structural information only on a single conformational state of the target protein. Thus, to obtain structural information on different conformational states, just one at a time, the protein has to be locked into a single conformation, which can sometimes be accomplished by mutagenesis, chemical crosslinking or addition of inhibitors or other effectors. These approaches usually work well for proteins that adopt a discrete number of somewhat stable conformational states. Structural heterogeneity that is caused by flexibly tethered domains, resulting in a continuum of different conformations, is substantially more difficult to deal with.

A distinct strength of single-particle EM is that it reveals not only the dominant conformation of the target protein but essentially all the conformations the protein can adopt. The reason is that images are taken of individual molecules, which can then be computationally sorted according to their appearance. While such classification procedures have been used early on to group single-particle images into different classes to calculate 2D class averages, more recently maximum likelihood-based 3D classification schemes have become practical and have been implemented in refinement software, such as RELION [18,19] and FREALIGN [20]. These programs use similar approaches and optimize not only the orientation parameters but also the ‘class’ or ‘occupancy’ parameter, yielding several density maps, which often reveal the target protein in different conformational states. A nice example in which 3D classification was used to reveal conformational changes is the EM study of glutamate receptors in the desensitized state. While 2D classification revealed that the desensitized α-amino-3-hydroxy-5-methyl-4-isoxazole propionic acid (AMPA) receptor adopts a continuum of conformations [21], 3D classification allowed visualization of the 3D structure of the receptor in some representative conformations [22] (Fig. 1).

Fig. 1.

Cryo-EM analysis of the conformational dynamics of AMPA receptors in the desensitized state. (a) Class averages of vitrified GluA2 homotetramers in the presence of fluorowillardiine, which stabilizes the receptor in the desensitized state [21]. The averages show that desensitization causes the two dimeric extracellular domains to separate, which then adopt a continuum of conformations. (b) Three 3D maps of GluA2 homotetramers in the desensitized state obtained by 3D classification. ATD, amino-terminal domain; LBD, ligand-binding domain; TMD, transmembrane domain. (b) Adapted from [22] and reprinted by permission from Macmillan Publishers Ltd: Nature, copyright 2014.

Single-particle EM allows visualization of larger assemblies

Many biological processes involve the interaction of binding partners that form stable or transient complexes. The understanding of their structural details is hampered by the challenges of crystallizing active machineries that are formed by multiple, sometimes only loosely associated proteins and that can exist in different functional and structural states. A further advantage of single-particle EM techniques is the ability to image large macromolecular machines at work, even for samples in which the multi-protein complex co-exists with unbound components. The potential of single-particle EM for visualizing large complexes involving membrane proteins is illustrated by the numerous studies of ribosomal complexes bound to the protein-conducting channel (SecYEG in prokaryotes; Sec61 in eukaryotes) [23–26].

Cautionary examples of membrane protein structures obtained by single-particle EM

Despite all the advantages of single-particle EM for protein structure determination, the technique has numerous pitfalls that can lead to incorrect 3D maps. The goal of single-particle EM is to generate a 3D density map from 2D projection images of a molecule at the highest possible resolution. This task is not trivial for numerous reasons: (i) it is not easy to determine the relative orientations of the EM projections, which is the basis for 3D reconstruction, especially because the images are very noisy; (ii) there may be more than one solution for reconstructing a 3D map from a given set of 2D images, especially if the images represent only a limited range of projection views; (iii) the images may contain molecules in different conformational states, which, if the images are not properly classified into different sets, can create a non-physiological structure that represents an artificial average of the molecule in different conformations, (iv) spurious features in an initial reference structure may be carried through the refinement procedure into the final density map, a problem known as reference bias; (v) structure refinement can lead to the alignment of noise, creating artificial features, a problem known as overfitting, and (vi) an easy way or criterion to assess whether the obtained 3D map is a faithful representation of the imaged molecule is not yet available. As a result, several structures obtained by single-particle EM, including some of membrane proteins, have been questionable or have eventually been found to be incorrect.

One example is work on the type I inositol 1,4,5-trisphosphate (IP₃) receptor, a 1.2-MDa ligand-gated ion channel that is responsible for Ca²⁺ release from the endoplasmic reticulum. From 2002 to 2004, five groups used different specimen-preparation techniques and image-processing approaches to produce density maps of the protein at resolutions ranging from 2 to 4 nm [27–31]. While the maps showed molecules of comparable size and overall shape as well as 4-fold symmetry (which was enforced during image processing), the maps differed considerably in domain arrangements and structural features, illustrating the problems that exist in producing a reliable low-resolution density map (Fig. 2a–d). While it was attempted to reconcile the different structures [33], a later single-particle cryo-EM study finally yielded a subnanometer-resolution map of the IP₃ receptor [32] (Fig. 2e). Acknowledging the controversial history, the authors invested significant effort to validate their structure [34].

Fig. 2.

Early, low-resolution density maps (a–d) and a more recent, intermediate-resolution density map (e) of the IP₃ receptor obtained by single-particle EM as seen from the cytoplasm (top row panels), parallel to the membrane (second row panels) and from the ER lumen (third row panels). The panels in the bottom row show central sections through the density maps at the position indicated by the lines in the panels in the third row, revealing the internal architecture of the receptor. (a) Density map at 24 Å resolution obtained with images of vitrified specimens collected at 120 kV on film, and the 3D reconstruction was calculated with the EMAN (and IMAGIC) software packages [29]. (b) Density map at 30 Å resolution obtained with images of vitrified specimens collected at 100 kV on film, and the 3D reconstruction was calculated with the EMAN software package [31]. (c) Density map at 24 Å resolution obtained with images of vitrified specimens collected at 300 kV on film, and the 3D reconstruction was calculated with the IMAGIC software package [30]. (d) Density map at 30 Å resolution obtained with images of negatively stained specimens collected at 200 kV on film, and the 3D reconstruction was calculated with the IMAGIC software package [27]. (e) Density map at 9.5 Å resolution obtained with images of vitrified specimens collected at 200 kV with a CCD camera, and the 3D reconstruction was calculated with the EMAN software package (EMD 5278) [32]. (a–d) Adapted and reprinted from Serysheva and Ludtke [33], Copyright 2010, with permission from Elsevier.

The AMPA receptor, a tetrameric glutamate receptor present in the post-synaptic membrane and responsible for fast excitatory synaptic transmission, is another example in which different groups obtained different single-particle EM structures, one with vitrified samples [35] and one with cryo-negatively stained preparations [36]. Only when the X-ray crystal structure was determined [37] was it possible to unambiguously identify the cryo-negative stain EM map as the more accurate representation of the receptor. Subsequent results on glutamate receptor biogenesis now indicate that the initial cryo-EM structure of the AMPA receptor may have been that of a dimeric assembly intermediate [38]. The cryo-negative stain EM study also showed that desensitization of the receptor results in the separation of the dimeric extramembranous domains, producing a Y-shaped molecule [36]. While a later cryo-electron tomography study of a related kainate receptor argued that the observed domain separation is due to staining artifacts [39], recent single-particle cryo-EM studies of vitrified AMPA receptors confirmed that receptor desensitization indeed results in the separation of the extramembranous domains and in a receptor that assumes a continuum of conformations [21,22].

Other examples in which single-particle cryo-EM produced questionable structures are the P2X receptor and HIV-1 envelope glycoprotein. In the case of P2X receptors, single-particle cryo-EM of the P2X₂ receptor produced a 15-Å density map [40] that turned out to be twice the size of the subsequently determined crystal structure of the closely related P2X₄ receptor [41]. In this case, it appears that it was the homemade software used to pick particles and process the images, which was not thoroughly tested and evaluated, that resulted in the incorrect density map. Another problem case is the HIV-1 envelope glycoprotein. Initial electron tomographic studies already produced differing structures. Investigators with little prior experience then produced single-particle cryo-EM density maps of the protein [42–44] that have been severely criticized by experts in the field [45–47]. These experts point out that the presence of particles is impossible to verify in the extremely noisy images and that the resulting structure suffers from a combination of model bias and overfitting.

New developments in single-particle EM (see below) have made it easier to obtain more reliable and higher resolution 3D maps. The risk of producing incorrect structures is thus much lower than it used to be, but only if the technique is used correctly. Given the increasing number of scientists that enter the field of single-particle EM and the remaining pitfalls of the technique, experts have begun to write primers to single-particle cryo-EM in an effort to guide new practitioners (e.g. [16,48,49]).

The recent quantum leap in single-particle cryo-EM

The recent introduction of direct electron detector device (DDD) cameras for imaging [50–52] ushered in a new era in single-particle cryo-EM [15,53,54]. The much better quality of the image data not only allows now for imaging of smaller molecules but also dramatically improves the effectiveness of image-processing steps, such as image alignment and 3D classification to computationally deal with specimen heterogeneity [18,20], thus making single-particle EM more broadly applicable to the structural study of membrane proteins.

DDD cameras

Image-recording devices are critical for cryo-EM data collection. Because only very low electron doses can be used to image biological samples in ice, it is essential that the detector records the image with excellent signal and noise performance, a characteristic described by the detective quantum efficiency (DQE), over a large spatial frequency range. Photographic film and scintillator-based digital cameras have been most widely used to record images in EM, but their DQEs drop quickly at higher spatial frequencies, i.e. they perform poorly for the recording of high-resolution information. Unlike scintillator-based cameras, DDD cameras directly register the incoming electrons and minimize backscattering, which results in a substantial improvement of the DQE, especially in the high-resolution range [55,56]. Detection of individual incident electrons, so-called electron counting, removes readout noise and variability in the electron signal, thereby further improving the DQE [51,57]. Electron counting improves the DQE at low resolution, which makes it possible to take images with a small defocus, so that high-resolution information is better preserved, while still providing sufficient image contrast for particle picking and alignment. Another advantage of DDD cameras is their high readout rates, which enable fractionation of a single exposure into a series of sub-frames, thus substituting the conventional one-second image with a multi-frame movie. Correction for the specimen movements that occur between these sub-frames then allows for computational compensation for motions resulting from the exposure of the specimen to the electron beam and instabilities of the specimen holder, thus minimizing image blurring and better preserving high-resolution information. Motion correction of movies collected with electron-counting DDD cameras now results in images of close-to-perfect quality [50,58] (Fig. 3). Furthermore, as motion correction makes images usable that previously would have had to be discarded, it greatly increases the yield of usable images and thus the efficiency of cryo-EM data collection.

Fig. 3.

The high quality of data that can be recorded with DDD cameras. (a, b) Average images of rotavirus double-layer particles in vitrified ice obtained from averaging 60 frames of a movie without aligning the frames to each other, showing substantial image blurring (a), and with aligning the frames, largely eliminating image blurring (b). (c and d) Single-particle cryo-EM density maps of transferrin–transferrin receptor complex (gold surface) with docked crystal structure shown as red ribbon diagram (PDB id 1SUV [59]). While the data for the two density maps were collected and processed in essentially the same way and the two density maps have similar resolutions of 7.5 and 7.0 Å, respectively, the map calculated with images recorded on film (c) [60] shows much less structural detail than the map calculated with images recorded with a DDD camera (unpublished data). (a) and (b) Reprinted from Brilot et al. [58], Copyright 2012, with permission from Elsevier.

3D Classification

Conformational variability of proteins has long been a major challenge for single-particle EM. In X-ray crystallography, the crystallization process essentially eliminates all major variations in protein conformation. In contrast, when a protein solution is quick-frozen during specimen preparation for cryo-EM, the vitrification process preserves all the conformations the protein can adopt. Hence, to produce high-resolution cryo-EM maps, it is critical to separate not only particles in different orientations into different groups but also particles that represent different conformations. 2D classification has been very useful to group particle images into homogeneous 2D classes. Classification is complicated by the low contrast of cryo-EM images, but it has been improved by the introduction of maximum likelihood-based algorithms that are particularly powerful for noisy data [61]. However, 2D images and class averages contain limited information, so 2D classification does not allow one to distinguish between classes that represent different 3D structures. This has to be done by 3D classification, which has recently been significantly improved by the development of effective maximum likelihood-based classification schemes [18,20] that now allow single-particle cryo-EM to use rather heterogeneous protein samples for structure determination. After sorting the particles into different 3D classes, higher-resolution density maps can be obtained by iteratively improving the orientation parameters, a process known as refinement. In the future, sophisticated 3D classification and refinement schemes may even make it possible to produce not only snapshots of a protein in different conformations but to generate movies that reveal – at high resolution – the continuous conformational change that underlies the function of a protein. However, accomplishing this goal will require very large data sets of high-quality images recorded with DDD cameras as well as sophisticated image-processing software. It will also greatly depend on the complete automation of the data collection and processing procedures [62].

The new DDD cameras and image-processing algorithms not only improve the achievable resolution of density maps but also make it possible to study smaller proteins. As a result, many more membrane proteins are now amenable to structure determination by single-particle cryo-EM. This was impressively demonstrated by the recent structures of transient receptor potential (TRP) channels and γ-secretase. The 4-fold symmetric, ∼300-kDa TRPV1 ion channel is the first membrane protein whose atomic structure was determined de novo by single-particle cryo-EM [63], and further studies on TRPV1 and TRPA1 provided first insights into the regulation of this important family of ion channels [64,65]. γ-Secretase is a membrane-embedded protease complex that is only 170 kDa in size and has no symmetry. Despite these challenges, it was possible to determine the structure of this protein at ∼4.5 Å by single-particle cryo-EM [66].

New ways to stabilize membrane proteins in aqueous solution

The primary challenge membrane proteins pose for structure determination is the hydrophobic surface of the transmembrane domain, which renders them unstable in aqueous solutions and promotes aggregation. Several approaches have been developed to stabilize membrane proteins in solution, which range from the more traditional detergents and small liposomes to the newer amphipols, nanodiscs and β-sheet peptides. Detergents are still the dominant reagents used to solubilize and purify membrane proteins, but new detergents are being developed that have better characteristics for structural studies, in particular very low CMCs. For example, maltose-neopentyl glycol-3 (MNG-3) analogs, which consist of two hydrophilic and two lipophilic groups that are linked by a central quaternary carbon [67,68], have been used in the X-ray crystallographic structure determination of several GPCRs. These detergents are likely to also be useful for single-particle cryo-EM, as their low CMC makes it possible to almost completely remove them from the protein solution, thus avoiding the problems encountered with traditional detergents, which lower the surface tension and make it difficult to obtain thin ice layers.

Amphipols, amphipathic polymers that interact with the transmembrane domains of membrane proteins in a semi-irreversible manner [69,70], have recently gained in popularity for the preparation of membrane proteins for cryo-EM, as they were successfully used in determining the high-resolution cryo-EM structures of TRPV1, γ-secretase and TRPA1 [63,65,66]. Similarly, inspired by the β-barrel proteins in the outer membrane of gram-negative bacteria, special β-sheet peptides have been designed and optimized to stabilize membrane proteins and preserve their activities in a membrane-mimetic environment [71]. β-Sheet peptides were successfully used to stabilize ATP-binding cassette (ABC) transporters and to study their conformational dynamics by negative-stain EM [71].

While very useful for the stabilization of membrane proteins, detergents, amphipols and β-sheet peptides are not perfect mimetics of a real lipid bilayer. For functional studies, membrane proteins are typically reconstituted into liposomes, which have also been explored for structural studies by single-particle cryo-EM. The first single-particle 3D reconstruction was obtained with a pore-forming toxin reconstituted into liposomes [72]. Although only at low resolution, the 3D map provided intriguing insights into the mechanism of pore formation. However, the heterogeneous size and strong lipid densities of liposomes pose great challenges to determining the orientation parameters of the reconstituted proteins. To overcome this issue, the random spherically constrained single-particle reconstruction method was developed [73] and then used to determine the structure of the human large-conductance calcium- and voltage-activated (BK) potassium channel reconstituted into liposomes [74].

While theoretically ideal, determining the structure of membrane proteins reconstituted into liposomes proved challenging. Nanodiscs are a new and exciting tool to determine the structure of membrane proteins in a true lipid bilayer environment. Nanodiscs are nanoscale discoidal lipid bilayers that are encircled and stabilized by a membrane-scaffold protein into which membrane proteins can be inserted [75–77]. Nanodiscs not only provide a native lipid bilayer environment for the reconstituted membrane protein, but they also allow for free choice of size and lipid composition. First cryo-EM structures of membrane proteins reconstituted into nanodiscs are those of SecY bound to the ribosome [24] and the RyR [78], which were determined to resolutions of ∼9 and ∼6 Å, respectively, and clearly resolved individual transmembrane helices, thus showing great promise for future studies.

Fab fragments as a tool to study small proteins

A known limitation for the application of single-particle EM analysis is that the target protein has to have a certain size. Theoretical considerations suggested that determination of a structure at atomic resolution requires the protein to have a molecular weight of at least 100 kDa [79]. However, in reality, proteins still need to be at least two or three times as big to allow structure determination at near-atomic resolution. This required minimum size, which is mainly due to the need for sufficient signal to align very noisy images with high precision, makes the majority of membrane proteins unsuitable for structure determination by single-particle EM.

Fab fragments, 50-kDa antibody fragments that retain the antibody's ability to specifically bind an epitope, have been introduced as a novel tool to make small proteins amenable to structure determination by single-particle cryo-EM. Fabs have been used in structural biology for many years and for various purposes, ranging from the stabilization of specific conformations and facilitating crystal packing in X-ray crystallography (e.g. [80,81]) to identifying specific domains in negative-stain EM studies (e.g. [36]). For structure determination by single-particle cryo-EM, Fabs provide several advantages [82]. In addition to simply increasing the size of the target protein, the bound Fab also provides a fiducial marker that helps in the computational alignment of the particle images. Fabs can also potentially stabilize a specific conformation of the target protein, and comparison of the density for the Fab in the final 3D reconstruction with the known crystal structure is a strong criterion to validate the density map.

To take advantage of Fabs for single-particle cryo-EM, the Fab has to be carefully selected. In particular, the Fab itself should be rigid, and its binding to the target protein has to (i) be specific, (ii) have a low off rate and (iii) occur in a rigid region of the protein, so that the relative orientation of the Fab to the target protein is always the same. Ideally, the Fab recognizes a conformational epitope rather than a linear sequence, as this can be used to increase the homogeneity of the data set [83].

Recombinant Fabs against specific epitopes can be produced easily by methods such as phage display [84]. The produced Fabs then need to be carefully analyzed by ELISA, size-exclusion chromatography and negative-stain EM, which should yield 2D averages in which the Fab can be clearly recognized and is always in the same position and orientation [82]. Once a suitable Fab has been identified, data collection and processing can proceed as normal. The initial work demonstrated that with the help of Fabs, single-particle cryo-EM can be used to analyze the structure of proteins as small as 65 kDa [82], and this approach later yielded a subnanometer-resolution density map of a heterodimeric ABC transporter [85] (Fig. 4).

Fig. 4.

The use of a Fab fragment for structure determination by single-particle cryo-EM. (a and b) Class averages (a) and 10-Å resolution density map (EMD 6087; [85]) (b) of the heterodimeric ABC transporter TmrAB in vitrified ice. (c) Two-dimensional class averages of the TmrAB transporter in complex with Fab AH5, clearly showing additional density for the Fab. (d) In the presence of the Fab, which provides additional mass and a fiducial marker for alignment, the resolution of the density map improved to 8.2 Å (EMD 6085; [85]). (a) and (c) Reprinted from [85] by permission from Macmillan Publishers Ltd: Nature, copyright 2015.

Membrane protein structures determined by single-particle EM

For many years, due to the required minimum size, only few large and stable membrane proteins were studied by single-particle EM. However, with the recent technical and methodological advances described above, many more membrane proteins became amenable to single-particle cryo-EM, which was powerfully demonstrated with the de novo structure determination of the TRPV1 channel [63]. This structure galvanized the structural biology community and ushered in an unprecedented flood of single-particle cryo-EM structures at subnanometer to near-atomic resolutions. In the following paragraphs, we briefly describe some of the membrane proteins that were studied by single-particle EM, from the early days until now.

RyR

RyRs are channels involved in the fast release of Ca²⁺ from the sarcoplasmic and endoplasmic reticula to the cytosol. These channels are activated by high cytosolic Ca²⁺ concentrations or by direct interaction with voltage-gated Ca²⁺ channels, and regulated by the binding of modulating proteins such as FKBP and calmodulin. The RyR homotetramers are the largest known ion channels, with a molecular mass of 2.2 MDa.

The history of structural studies on RyRs closely parallels the development of single-particle EM. The first 3D density maps appeared in the mid-1990s and had resolutions between 20 and 30 Å [86,87]. It took 10 more years and technical and methodological advances to push the resolution of the map to about 10 Å, which allowed a better understanding of the large cytoplasmic domain of the receptor that accounts for about 80% of the protein mass [88]. The structure of the receptor in the closed state was followed by structures at higher resolution and of the open state, providing first insights into the gating mechanism [89,90]. Recently, with the use of DDD cameras and 3D classification, three new density maps were published with resolutions up to 3.8 Å (Fig. 5a). These studies allowed atomic models to be built [91,95] and confirmed the previously proposed gating mechanism [78].

Fig. 5.

Examples of membrane protein structures determined by single-particle cryo-EM. Density maps are shown as gold surfaces and atomic models as red ribbon diagrams. The black lines indicate the approximate position of the membrane. (a) Density map of the 2.2-MDa homotetrameric ryanodine receptor at 4.8 Å resolution (EMD 6107, PDB 3J8E; [91]). (b) Density maps of V-type ATPases obtained from data collected with a CCD camera at 9.7 Å (EMD 5335, PDB 3J0J; [92]) (left panel) and from data collected with a DDD camera at 6.9 Å (EMD 6284, PDB 3J9 T; [93]) (right panel). (c) Density map of respiratory complex I at 4.95 Å resolution (EMD 2676, PDB 4UQ8; [94]). (d) Density map of the 300-kDa homotetrameric TRPV1 channel at 3.4 Å resolution (EMD 5778, PDB 3J5P; [63]).

ATPases

ATP synthases and ATPases, which use the membrane potential to synthesize ATP or use the energy released from ATP hydrolysis to pump ions across membranes, respectively, are another family of membrane proteins that has been studied extensively by single-particle EM [96]. The density map of the ∼600 kDa vacuolar-type ATPase (V-ATPase) from a thermophilic bacterium was the first single-particle cryo-EM structure of a membrane protein that was determined to subnanometer resolution, and the map clearly resolved some of the transmembrane α-helices [92] (Fig. 5b, left panel). With the recent technical advances, higher resolution maps have now been determined, namely that of the mitochondrial ATP synthase dimer from the colorless green alga Polytomella sp. at 6.2 Å, which provided insights into the coupling between proton translocation and ATP production [97], and three structures of the V-ATPase from Saccharomyces cerevisiae, at resolutions from 6.9 to 8.3 Å, that show the ATPase in different rotational states, thus revealing for the first time the coordinated conformational changes of the rotor and stator during its enzymatic cycle [93] (Fig. 5b, right panel). These exciting structures are likely just the first of many more high-resolution single-particle cryo-EM structures of ATP synthases and ATPases in different functional states that will provide further insights into the energy transduction pathway during ATP synthesis and hydrolysis.

Respiratory complex I

The enzymes of the respiratory chain are among the most studied complexes, not only because of their biological importance but also because of their large size and relative abundance in natural sources. While the structure of all other respiratory complexes were determined by X-ray crystallography, the NADH:ubiquinone oxidoreductase (complex I) proved refractory to crystallization attempts. Obtaining structural information for complex I therefore depended on single-particle cryo-EM, and the study of this specimen indeed prompted the development of FREALIGN [98]. As with other membrane proteins, the recent technological advances led to a substantially better EM map of complex I from bovine heart mitochondria with a resolution of 5 Å. The map allowed structural models to be built for 14 conserved core subunits and 14 supernumerary subunits, providing new insights into the regulation, assembly and homeostasis of complex I [94] (Fig. 5c).

TRP channels

TRP channels are nonselective cation channels that are present from yeast to humans and respond to a wide variety of stimuli, including temperature, taste, pH, physical force and light [99]. Although TRP channels display only low sensitivity to changes in membrane potential, previous studies suggested structural similarity with voltage-gated K⁺ and Na⁺ channels [100,101]. Because of their biological importance, substantial efforts have been invested in obtaining structural information for this family of channels by X-ray crystallography, but to little avail.

In late 2013, taking advantage of the recent technical advances, a single-particle cryo-EM reconstruction of the TRPV1 ion channel at 3.4 Å resolution was published. The map showed clear side-chain densities and thus allowed de novo building of an atomic model for the entire channel, becoming the first atomic structure of a membrane protein determined by single-particle EM [63,64] (Fig. 5d). The atomic structure of TRPV1 confirmed the similarity in the general transmembrane domain architecture between TRP and voltage-gated channels, but also highlighted the structural differences that underlie their unique characteristics in function and regulation [63]. While the structure of TRPV1 in the apo state revealed a closed channel, two more cryo-EM structures of TRPV1 with specific agonists and modulator bound showed the channel in different open conformations, thus revealing the mechanism of channel activation by ligand binding [64]. A recent cryo-EM structure of the TRPA1 channel provided further insights into distinct regulatory mechanisms for different TRP channels [65].

The importance of the TRPV1 structure for single-particle cryo-EM cannot be overstated, as it showcases the power of this approach not only to determine novel protein structures but also to provide deep mechanistic insights into their function. The TRPV1 structures were made possible by a combination of several technical developments, including the use of amphipols to stabilize the protein, imaging with an electron-counting DDD camera with dramatically enhanced DQE, sub-frame alignment to correct for motion-induced image blurring, and 3D classification implemented in RELION [19]. This landmark study made it clear that single-particle cryo-EM can finally deliver on its long-held promise of being able to generate atomic structures without the need of growing crystals. In fact, the TRPV1 structures were the beginning of an avalanche of high-resolution cryo-EM structures of both soluble and membrane proteins.

Potassium channels

Structural work on K⁺ channels has been dominated by X-ray crystallography, which yielded numerous structures and important biological insights, including a first structural view of voltage gating [80]. Nevertheless, single-particle EM contributed to the identification of the conformational rearrangements of the voltage sensors that enable gating in response to changes in the membrane potential [83]. In this study, high-affinity Fabs were used to stabilize the channel in the open conformation, but the Fabs also increased the size of this small channel of only ∼100 kDa and made it amenable to cryo-EM.

In a different study, aimed at visualizing the voltage sensor in a membrane environment, a large-conductance Ca²⁺- and voltage-activated K⁺ channel was reconstituted into small liposomes. After imaging the liposomes, single-particle reconstruction took advantage of the constant relative orientation of the protein with respect to the liposome surface, thus simplifying the determination of the orientation parameters for each particle [74]. While this study demonstrated that the membrane curvature does not affect the overall conformation of the protein, at a resolution of 17 Å the density map did not allow the authors to establish whether the lipid environment caused small changes in the conformation of the transmembrane domain.

More recently, the structure of the Slo2.2 channel was determined to a resolution of ∼4.2 Å, which made it possible to build a model of the channel [102]. The structure of the closed channel revealed commonalities and differences to the structure of voltage-gated K⁺ channels as well as the structural basis for the very fast K⁺ conduction of this channel. A structure of the open channel will provide further insights into channel gating.

ABC transporters

ABC transporters harness the energy of ATP binding and hydrolysis to translocate a large variety of substrates across lipid membranes. While available X-ray structures of ABC transporters in inside- and outside-facing conformations are consistent with the alternating-access model for substrate transport [103], the detailed mechanisms underlying conformational changes and substrate translocation remain under debate, mainly due to the lack of structures that represent intermediate conformations. ABC transporters have generally been considered to be too small for single-particle cryo-EM studies. However, taking advantage of a tightly binding Fab, a recent study on detergent-solubilized TmrAB, an ABC transporter thought to be involved in multidrug resistance, generated a cryo-EM map of the TmrAB-Fab complex at 8.2 Å resolution that clearly resolved the transmembrane helices of the transporter [85]. Circumventing the issue of crystal contacts, single-particle EM has the potential to visualize the conformational dynamics of unconstrained ABC transporters and to provide a complete view of the conformational changes that underlie the transport cycle of ABC transporters. Toward this goal, a recent negative-stain EM study on bacterial MsbA and mammalian P-glycoprotein, both stabilized by β-sheet peptides, delineated the entire conformational spectrum of these two transporters during ATP hydrolysis [104].

Pore-forming toxins

Pore-forming toxins (PFTs) are specialized proteins that evolved to punch holes into target cell membranes. PFTs are usually secreted as water-soluble monomers that oligomerize on the target membrane into so-called pre-pores that then undergo a conformational change that leads to the formation of a transmembrane pore [72]. Single-particle cryo-EM produced several structures of PFTs in the pre-pore and pore states that revealed a common mechanism involving the extrusion of a bent β-sheet to form a transmembrane β-barrel pore and elucidated the potential trajectory from domain buckling to transmembrane region insertion [105,106]. A recent 2.9 Å resolution cryo-EM structure of the pore formed by the anthrax protective antigen (PA) revealed the organization of the membrane-spanning translocation channel and provided support for the Brownian ratchet model for protein translocation [107]. The structure also revealed the catalytic Φ-clamp, which is composed of phenylalanine residues and is critical for efficient toxin translocation while preventing the passage of cations. However, a full understanding of protein translocation will need a structure of the PA pore in the process of transporting a polypeptide.

Concluding remarks and outlook

Almost from its inception, single-particle cryo-EM has been used to study the structure of membrane proteins, but only the introduction of DDD cameras has made it possible to use this approach to determine their structure at atomic resolution. While single-particle cryo-EM has the enormous advantage that it allows structure determination without the growth of crystals, many membrane proteins remain too small to be studied by single-particle cryo-EM. Fab binding can currently be used to bring some membrane proteins into the amenable size range, but in future, DDD cameras with even better DQEs and possibly the use of phase plates may allow the study of increasingly smaller membrane proteins. Until then, however, small membrane proteins will largely remain targets for structure determination by X-ray crystallography. Single-particle cryo-EM does not always produce density maps at a resolution suitable to build an atomic model, but now usually delivers maps at subnanometer resolution. At this resolution, α-helices are clearly resolved, making flexible fitting of crystal structures (or homology models) into the map very powerful so that at least most of the main chain can be traced. In addition, single-particle cryo-EM provides information that cannot be obtained by X-ray crystallography, such as information on protein dynamics, provided by 3D classification, and information on the effect of a true lipid bilayer environment on the protein structure, provided by the use of nanodiscs. Hence, while single-particle cryo-EM will not replace X-ray crystallography, it has become a very powerful and versatile tool to study the structure of membrane proteins with the future holding much promise for playing an even bigger role in the structural studies of membrane proteins.

Funding

Funding was provided by the Howard Hughes Medical Institute (to T.W.).