Sample preparation is the first and most important step in cryo-EM because the purity and quality of the sample directly affect the cryo-EM map quality (Passmore and Russo, 2016). In Section 3.1, three applicable techniques are described in the sample preparation step: 1) sample condition screening, 2) techniques for membrane proteins, and 3) techniques for small (<50 kDa) proteins.

Owing to limited instrument accessibility and time-consuming nature of ensuring good data quality in cryo-EM, it is necessary to evaluate the sample quality before collecting and analyzing cryo-EM data. Various factors influence the quality of single particles, such as pH, salt concentration, storage conditions, sample concentration, and purification steps (e.g., size exclusion chromatography) (Kampjut et al., 2021). For this reason, classical methods such as gel electrophoresis, size exclusion chromatography and dynamic light scattering (DLS) are required to scan and evaluate sample quality at the points of purity, sample heterogeneity of oligomeric states, sample aggregation and stability of complex state. Additionally, the analysis of sample quality using negative staining transmission electron microscopy (TEM) is a powerful and fast tool to estimate the adsorbed conditions of samples on the grids visually for parameters such as particle distribution (Scarff et al., 2018; Gonen, 2021).

Membrane proteins comprise a protein class that has benefited most from improvements in cryo-EM. To stabilize the membrane proteins, detergents and membrane mimetics are required. Autzen et al., 2019 mentioned various approaches in their review. One approach, introduced in the review, involves using detergents such as n-dodecyl-β-D-maltopyranoside, maltose neopentyl glycols, digitonin and glycol-diosgenin. Cholesterol derivatives, including cholesteryl hemisuccinate and CHAPSO, are also widely used in studies on membrane proteins. Another membrane-mimetic approach involves the use of amphipols, short hydrophilic protein polymers with hydrophobic side chains that cover the hydrophobic sites of membrane proteins (Tribet et al., 1996). The membrane proteins with amphipols are free from the thick belt near the transmembrane domain formed by detergents and nanodisc, one of the obstacles to determine membrane protein structure (Venien-Bryan and Fernandes, 2023). In that point, using amphipol is beneficial in the refinement process, the last stage of structure determination using cryo-EM. In addition to the detergents and amphipols, membrane scaffold proteins have been used to stabilize lipid bilayers by forming various lipid nanodiscs with membrane proteins (Denisov and Sligar, 2016; Efremov et al., 2017). Nanodiscs are prepared as follows: 1) incubation of detergent-solubilized proteins, lipids, and membrane scaffold proteins and 2) removal of detergents. In the procedure, selection of appropriate membrane scaffold proteins is critical to mimic the physiological status of membrane proteins. In 2019, Ognjenovic et al. published a review in which almost 50 structures of membrane proteins including ion channels, transporters, receptors, and others were determined using detergents, amphipols, and nanodiscs at a resolution above 4 Å (Ognjenovic et al., 2019). One of the recent cases with high resolution is the structure of the SARS-CoV-2 3a, a non-selective cation channel and a regulator of viral pathogenesis, determined at 2.1 Å resolution using lipid nanodiscs (Kern et al., 2021).

Although many advancements have been made in the field of structural biology in recent years, small protein structures remain challenging targets because of their inadequately distinguishable structural characteristics and low signal-to-noise ratios (Yeates et al., 2020; Zhang et al., 2022). By 2 August 2023, 9,872 EM maps released in the EMDB were from targets over 100 kDa, but only 117 (≈1.1% of the total) were from targets below 50 kDa. The 117 EM maps included models that were obtained by focused refinement of large molecules; the number of targets below 50 kDa was lower. To overcome this challenge, the strategy of increasing the molecular weight of target proteins is widely applied using scaffolds composed of an adaptor-specific scaffold core (or platform base) and a target-specific adaptor (Yeates et al., 2020). Two approaches have been reported for the development of adaptors: 1) designing a selective target-binding adaptor and 2) designing a fusion protein of an adaptor and target protein (Wentinck et al., 2022). Target-binding adaptors include antibodies and small antibody fragment-based proteins, such as antigen-binding fragments (Fabs) and nanobodies. The development of a fusion protein containing an epitope that the known antibody/nanobody/Fab recognizes with high specificity is another approach. By introducing the sequence from apocytochrome b₅₆₂RIL (BRIL, ≈11 kDa) (Mukherjee et al., 2020; Tsutsumi et al., 2020), the glycogen synthase domain of Pyrococcus abyssi (PGS, ≈20 kDa) (Zhang et al., 2022) and the third intracellular loop from the kappa opioid receptor (κOR-ICL3) (Robertson et al., 2022b) to GPCRs, the target structures were determined using their specific nanobodies and Fabs (Wentinck et al., 2022; Zhang et al., 2022). The inserted regions such as BRIL, PGS, and κOR-ICL3 were developed as crystal chaperones, and also have been used for structure determination by cryo-EM (Chun et al., 2012; Mukherjee et al., 2020; Zhang et al., 2022; Miyagi et al., 2023). For instance, the BRIL sequence was inserted into the N-terminal region of solute carrier family 19 member 1 (SLC19A1) (65 kDa) with its substrate 5-methyltetrahydrofolate (5-MTHF) (Dang et al., 2022). The structure of the BRIL-containing SLC19A1–5-MTHF complex was determined at 3.5 Å resolution. In addition to adaptor proteins, scaffold cores, such as glutamine synthetase (dodecamer, D6), have been applied together with adaptors to increase the size by oligomerization of adaptor–target complexes (Coscia et al., 2016).

After obtaining high-quality purified specimens, grid optimization is required to obtain vitrified samples on an EM grid with an appropriate ice layer thickness, the most time-consuming process (Benjin and Ling, 2020; Xu and Dang, 2022). Generally, on the gold or copper grid containing open holes, 3 µL of a purified sample is placed to form a thin layer ideally <100 nm thickness in the holes (Chua et al., 2022). The excess sample on the grid surface is then gently blotted using filter paper, and the grid is rapidly frozen in liquid propane or ethane cooled with liquid nitrogen for vitrification (Liu and Wang, 2023). Following this procedure, the optimal single-particle grid for cryo-EM is prepared as a single-particle state in the holes but away from the air–water interface (AWI). The adsorption of single particles on the AWI results in an orientation bias and/or partial or full denaturation of the drug targets by exposure of less hydrophilic sites to the AWI (Klebl et al., 2020).

In the grid preparation, various factors affect a single-particle grid, such as the grid material, glow discharge process, incubation conditions of proteins, and blotting protocols (Weissenberger et al., 2021). After blotting, the residual liquid layer on the grid bar or film is recruited, which affects the sample behavior (Glaeser, 2018). In some cases, these factors lead to disassembly, denaturation, and aggregation of macromolecules, particularly proteins, during sample preparation on the grid (Drulyte et al., 2018).

The ice thickness and AWI of single-particle cryo-EM grids are crucial in two respects: the orientation and the overlap of particles (Noble et al., 2018). Ice, which is thicker than the major axial length of a single particle or the minor axial length with at least an additional space of 20 nm, is required for the random orientation of particles. Otherwise, the particles are aligned near the AWI with a biased orientation distribution. In contrast, sufficiently thin ice is required to eliminate the overlap of multiple particles in the beam direction. Several methods have been developed to prepare ideal single-particle cryo-EM grids, such as the use of additives to form a protective layer during AWI (Chen et al., 2019), on-grid supports (Han et al., 2020), and vitrification devices to shorten the freezing time (Ravelli et al., 2020).

First, amphipathic molecules are the most common additives that naturally block AWI and enable the formation of randomly oriented particles. For example, CHAPSO, a zwitterionic detergent, solves problems related to the destabilization, aggregation, and/or preferential orientation of most specimens (Chen et al., 2022). For example, Ye et al., 2022 solved the structure of SARS-CoV-2 omicron spike protein ectodomain at 3 Å resolution using, but not being limited to, CHAPSO. They showed that the trimeric receptor binding domains (RBDs) of omicron spike proteins were mainly in the open (“standing up”) conformation, one of three RBDs up and the others down, as ready state for receptor binding.

Second, to overcome the AWI issue, the use of grids with carbon supports is suggested as another strategy in the grid preparation step to enhance image quality. Generally, two major supports are used: amorphous carbon-based and graphene-based supports (Han et al., 2020; Glaeser, 2021; Patel et al., 2021). Coating method with continuous thin films on grids can provide separate interaction surfaces, rather than AWI in “open hole” grids, for the target molecules. Instead of holes, the particles are distributed on the supports coated on a copper or gold grid as the film-water interface. Amorphous carbon films are commonly used to coat grids of large macromolecules (>300 kDa) (Grassucci et al., 2007). The amorphous carbon films are typically the first choice for challenging samples and may cause notable levels of background noise during imaging (Russo and Passmore, 2014). Graphene is a thin 2D nanomaterial that has been demonstrated to be an effective solution to reduce background noise, uneven particle distribution, and ice thickness (Naydenova et al., 2019a; Han et al., 2020).

The development of the vitrification process began in the 1980s with water (Dubochet and McDowall, 1981) (Figure 2). Using vitrification devices, rapid freezing (on the scale of millisecond or less) is expected to minimize the interaction of molecules with AWI by reducing the diffusion of particles to inhibit AWI (Klebl et al., 2020). Vitrobot (Thermo Fisher Scientific) and GP2 (Leica Microsystems) are commonly used plunge freezers. VitroJet, developed by Ravelli et al., 2020, enables the vitrification time to decrease to as low as 80 ms by automating grid preparation by placing samples on grids for freezing. VitroJet requires sub-nanoliter amounts of sample to be applied onto a grid using the pin-printing method, which is then immediately cooled with a cryogen, eliminating the blotting step. Similar to the VitroJet, the Spotiton is another grid-preparation device that has been shown to decrease sample vitrification time to approximately 10 ms by combining voltage-assisted spraying and vitrification (Jain et al., 2012; Razinkov et al., 2016; Kontziampasis et al., 2019).

The next stage in the cryo-EM workflow is data collection using a cryogenic TEM (cryo-TEM). Electron microscopes are composed of five main parts: 1) an electron-beam-producing source, 2) a group of magnetic lenses, 3) a vacuum system, 4) a cryogenic sample holder, and 5) a detector for image capture (Chua et al., 2022). Since their introduction into the cryo-EM field, these microscope components have undergone major changes and have been developed to enable the collection of high-resolution structural data (Figure 2).

With the development of EMs, electron energy sources have improved to enhance the spatial and temporal coherence of electron beams from the initial energy sources such as tungsten filaments and LaB₆ crystals (Joy, 2019). As LaB₆ cathodes have narrower tips than tungsten filaments, they generate smaller radius electron beams with better coherence (Tang et al., 2021). Subsequently, the advanced electron guns, field-emission guns (FEGs) were developed in scanning electron microscope (Figure 2) (Crewe et al., 1968; Tonomura, 2011). Then, TEMs equipped with cold FEGs (CFEGs) were developed (Ricolleau et al., 2012). The CFEGs produced electron beams with an energy spread of approximately 0.3 eV, maintaining beam brightness when operating near room temperature (Hamaguchi et al., 2019). This energy spread is almost half that of the general FEG, resulting in a better signal-to-noise ratio owing to the improved coherence of the electron beam (Ricolleau et al., 2012; Kato et al., 2019).

With the development of the CFEGs, an ultrahigh vacuum (10⁻⁸–10^–9 Pa) system is required for cryo-EM (El-Gomati et al., 2021). The ultrahigh vacuum system removes vaporized water from the vitrified sample and contaminants near the tip and prevents interference from the electron beam and induces electron diffraction by the contaminants and water molecules (Kogure, 2013; Cheng, 2015).

A phase plate introduces a phase shift in the diffraction plane of a microscope, resulting in phase contrast (Danev and Baumeister, 2016). Therefore, it improves the signal-to-noise ratio by providing enhanced image contrast and in-focus data acquisition (Glaeser, 2013). The phase plates have been used in light microscopes for a long time but were unavailable for use in electron microscopes until recently (Danev and Baumeister, 2016). The Zernike phase plate, a thin material film with a central hole and a phase shift of ∼π/2, has been used in cryo-EMs, although it has the disadvantages of a short lifespan and fringe artifact creation (Danev et al., 2009; Glaeser, 2013). Moreover, the difficulty in aligning the beam center to a small hole is a barrier to the use of Zernike phase plate. Therefore, a new phase plate, the Volta phase plate (VPP), was introduced in the field of EM for biological samples in 2014 (Danev et al., 2014). The VPP is a continuous amorphous carbon film with working conditions of approximately 25 mA and 200°C to prevent contamination. It has no fringe artifacts, a long lifespan, and no alignment requirements. Using the VPP, the full-length calcitonin receptor (CTR), a class B GPCR was reconstructed in a complex with peptide agonist and Gαβγ heterotrimeric protein, as a therapeutic target for several bone diseases (Liang et al., 2017). In addition to thin film-based phase plates such as Zernike phase plate and VPP, two new types of phase plates, magnetic phase plates using a magnetic field and laser phase plates using photon–electron interactions, have been developed (Wang and Fan, 2019). Very recently, the laser phase plates was applied to cryo-EM; however, it has not yet been commercialized (Axelrod et al., 2023).

Aberration correctors are another critical component for improving the power of cryo-EM for high-resolution structures. (Evans et al., 2008). There are two major aberrations: spherical aberration caused by failure of convergence of the paraxial ray and marginal ray passing through the objective lens, and chromatic aberration caused by various wavelengths of the electron beam. The spherical aberration was removed using correctors and aspherical lenses (Fan et al., 2017). To remove chromatic aberration with a FEG and a monochromator, a chromatic aberration correctors composed of an electromagnetic hexapole or quadra/octa-pole was also applied (Leary and Brydson, 2011). Eliminating these aberrations using correctors enhances the image quality of the focused beam (Freitag et al., 2005).

Photographic films, charge-coupled device (CCD) detectors, and complementary metal oxide semiconductors (CMOSs) were routinely used for image detection and recording in EMs. Photographic films have the advantage of collecting images of a large size, but require additional processes, such as digitizing the acquired images for further processing and analysis. Compared with films, CCD and CMOS detectors offer automated data acquisition and immediate image access for analysis. Usually, to collect images in these detectors, a single instant electron is scattered by scintillators, releasing photons. The photons passing through the optical fibers are then detected by a CCD or CMOS. In this procedure, photons generated from scattered single instant electrons reduce the resolution of spatial information and decrease signal-to-noise levels by noise generation (Meyer and Kirkland, 1998). Additionally, the image quality obtained from detectors coupled with scintillators and fiber optic plates is appropriate at 100 keV, whereas the noise increases at high powers, such as 200 and 300 keV, which are the general energies of cryo-EM (Bammes et al., 2012; McMullan et al., 2014). Although CMOS detectors have advantages over CCD detectors, such as faster frame rates and lower blurriness, problems caused by using scintillators and optic fibers remain. A new technique, the direct electron detector (DEDs), was designed in 2004 and applied to cryo-EM (Xuong et al., 2004; Bai et al., 2015). There are two types of DEDs: 1) monolithic active pixel sensors (MAPS) (Peric, 2007) and 2) hybrid pixel array detectors, including Medipix sensor-based detectors (Jakubek et al., 2004) and electron microscope pixel array detectors (Tate et al., 2016). Among these, MAPS exhibited the best performance at high electron beam energies. Therefore, MAPS are widely used in cryo-EM at 200 and 300 keV. In DEDs, unlike previous detectors, MAPS rather than scintillators and optic fibers, transfer electrons to detectors directly, and DEDs enable the detection of electrons with high resolution and high signal-to-noise ratio by reducing noise and reducing the loss of signal in the electron-counting mode (Levin, 2021). Moreover, DEDs exhibit a high frame rate when collecting images. This improvement in detectors leads to the resolution revolution in cryo-EM.

Using high-voltage (300 keV) cryo-TEMs routinely, protein structures were determined at high resolution. With other technical advancements, including FEGs and DEDs, the low-voltage microscopes such as Glacios (Thermo Fisher Scientific, MA, United States) and CRYO ARM 200 (JEOL Ltd., Japan) are also used for structure determination (Merk et al., 2020; Wu et al., 2020; Koh et al., 2022; Thangaratnarajah et al., 2022). Various studies have shown membrane protein structures at 2–4 Å resolution using 200 keV microscopes (Fan et al., 2020; Oosterheert and Gros, 2020; Cao et al., 2021; Budiardjo et al., 2022; Zarkadas et al., 2022). As an example, the inhibited and the active states of human mitochondrial calcium uniporter (MCU) holocomplex structure, comprised of four main proteins: MCU, mitochondrial calcium uptake 1 and 2, and an essential MCU regulator were determined at 3.6 Å resolution (Fan et al., 2020). The more advanced 100 keV cryo-EM was also developed and applied to determine several structures with diverse sizes in 2019 (Naydenova et al., 2019b). They reported structures of macromolecules in the range of 64 kDa–4.5 MDa size within 3.4–8.4 Å resolutions. Later, using the 100 keV cryo-EM, macromolecules with size range between 140 kDa and 2 MDa such as 70 S ribosome and GABA_A receptor were structurally determined within 2.7–4.5 Å resolution range (McMullan et al., 2023). The development of low energy cryo-EM enhanced the user accessibility to cryo-EM while increasing the number of available cryo-EM and reducing costs associated with the equipment and its installation and maintenance (A low-cost electron microscope maps proteins at speed, 2023). Thus, it is expected that the lowered accessibility barriers will accelerate determination of targets structures for SBDD.

Cryo-EM is differentiated from X-ray crystallography by using “images” instead of diffraction patterns as primary data (Wang and Wang, 2017). A typical image-processing workflow for cryo-EM data can be divided into five steps: 1) data pre-processing, 2) particle picking and extraction, 3) 2D classification, 4) 3D reconstruction, and lastly 5) 3D refinement (Lyumkis, 2019; Chung et al., 2022). After releasing SPIDER and WEB in 1996 (Frank et al., 1996), several software packages are available for the processing steps, including EMAN (Ludtke et al., 1999; Tang et al., 2007; Bell et al., 2016), RELION (Scheres, 2012; Zivanov et al., 2018), cryoSPARC (Punjani et al., 2017), and cisTEM (Grant et al., 2018).

After collecting the images, data pre-processing is initiated by correcting the images of particles moved by electron beam exposure. Most particle movement occurs for a very short time in the early period of beam exposure, after which the movement decreases spatially and temporally (called stable but not fixed) (Li et al., 2013). Therefore, rejection of the images obtained during the early period was suggested as a solution to remove the blurriness of the images due to beam-induced particle motion. However, this strategy was not always welcome because the longer the exposure to the sample, the more radiation damage accumulated in the samples, as shown in Table 1. This dilemma has been solved through the development and advancement of DEDs, as mentioned in Section 3.3. The high frame rate (10–400 frames/s) of DEDs allows images to be captured as multi-frames, such as a movie, in a single exposure. Although it produces a low signal and low-resolution contrast by short exposure at the millisecond level, high-resolution images can be obtained. Moreover, a high frame rate enables the collection of images that are stable and undamaged by beam exposure (Li et al., 2013; Zivanov et al., 2019). Therefore, using images collected from the least damaged particles in the early frames is more advantageous despite the blurriness caused by the high movement of particles (Cheng et al., 2015). Software has been developed in addition to cameras. Software such as MotionCor2 (Zheng et al., 2017) and alignframes_lmbfgs (Rubinstein and Brubaker, 2015) were developed to estimate and correct frame motion using motion correction algorithms, resulting in motion-corrected images. To assess the quality of the micrographs, contrast transfer function (CTF) estimation is performed. CTF obtained using software such as CTFFIND4 (Rohou and Grigorieff, 2015) and Gctf (Zhang, 2016) is required to curate micrographs based on the estimated defocus and astigmatism levels.

The particle-picking step to locate the target molecules in micrographs is challenging for several reasons, such as low signal-to-noise ratios, impurities on the micrograph, non-uniform distributions, preferential orientations, and undistinguishable structural characteristics (Chua et al., 2022). Several groups suggested the required number of collected particles from micrographs to obtain high resolution EM maps, as “the more, the higher.” In 2016, Danev and Baumeister reported 5,000–10,000 particles were required to achieve below 3.5 Å resolution with Fourier shell correlation (FSC) = 0.5 criterion (Danev and Baumeister, 2016). However, Chua and the colleagues suggested that ≈1,000,000 particles were required to obtain structures at 3.5 Å or higher resolution with FSC = 0.143 criterion (Chua et al., 2022). They estimated the total number of requested micrographs as 5,000–10,000, which was acquired from division of ≈1,000,000 particles by the average number of analyzable particles per micrograph, ≈100 particles/micrograph.

The initial software for particle picking used user intervention, which was a time-consuming approach. The semi-automated pickers were developed based mainly on template-base method (applied into the particle picker in RELION-1.3 (Scheres, 2012; 2015)), feature-based method (applied to DoG Picker in Appion (using difference of Gaussian) (Voss et al., 2009), and Laplacian of Gaussian-applied picker in RELION (Kimanius et al., 2021) and Blob Picker in cryoSPARC (Punjani et al., 2017; Chung et al., 2022; Wu et al., 2022). The results obtained using semi-automated pickers were more efficient than those obtained using manual pickers, but the selected input images or features induced a bias. To avoid this bias, the convolutional neural networks-based particle pickers has been raised such as DeepPicker (Wang et al., 2016), DeepEM (Zhu et al., 2017), Topaz (Bepler et al., 2020), Warp (Tegunov and Cramer, 2019), crYOLO (Wagner et al., 2019; Wagner and Raunser, 2020), PIXer (Zhang et al., 2019), and DeepCryoPicker (Al-Azzawi et al., 2020). Their algorithms are composed of two steps: 1) a training network using manually selected sets of micrographs, and 2) picking particles automatically using trained algorithms.

In the third step, 2D classification is the process of discriminating and refining clearly aligned images of particles from other images, and grouping the particles based on conformation and composition for 3D classification. To cluster the similar particle images in the 2D classification step, the maximum-likelihood (Sigworth, 1998) or cross-correlation (CC) (Radermacher and Ruiz, 2019) approach is applied to align and average particles (Figure 2). Using the CC approach, images are aligned to acquire the maximum CC coefficient value between two images or between a single image and the average of the images. However, the CC approaches can generate high but false correlation coefficients, particularly in the images with low signal-to-noise ratios, which are usually obtained from the DED images. Unlike CC approaches, the maximum-likelihood approach calculates the probability-weighted averages of all possible orientations of images (Sigworth et al., 2010). With the hidden variables assumption, the maximum-likelihood algorithm has been broadly applied to software packages such as FREALIGN (Grigorieff, 2007), RELION (Scheres, 2012), and cryoSPARC (Punjani et al., 2017). The representative images acquired after 2D classification are again used as a training set for particle picking, and then picking particles and 2D classification followed as an iterative process to sort particles more thoroughly into a set of high-quality particles (McSweeney et al., 2020).

Three steps are required to obtain optimally estimated 3D EM maps from 2D particle images: 3D reconstruction, 3D classification, and refinement. Therefore, the three steps are integrated as described in the previous paragraph. Three-dimensional construction is performed using a known reference structure or Ab initio model (Joubert and Habeck, 2015). References can be obtained from databases (PDB and EMDB) or Random Conical Tilt-applied 3D models from negative-stained EM images. Ab initio models are constructed using various algorithms, such as the common-lines-based model, random-model methods, stochastic hill climbing, stochastic gradient descent, and Bayesian approach (Joubert and Habeck, 2015). For example, the stochastic hill climbing algorithm was used in SIMPLE/PRIME (Reboul et al., 2016; Reboul et al., 2018) and the stochastic gradient descent algorithm was implemented in the 3D reconstruction step in cryoSPARC (Punjani et al., 2017) and RELION (Scheres, 2012). One or more initial models are constructed from the reference structure by using a projection-matching algorithm (Nogales and Scheres, 2015). Using the initial models, an unbiased 3D classification is performed, consequently helping to identify multiple conformations or separate junk particles. Although the process of 3D classification is the same as that of 2D classification, it is critical to distinguish between the various conformations of the target molecules or the different complex structures with subtle differences (Murshudov, 2016).

Three-dimensional refinement is the final step in refining a 3D EM map to a high resolution, finding the optimal orientation for 2D particles using the initially reconstructed map in the previous step as a reference. The branch-and-bound algorithm (Lawler and Wood, 1966; Punjani et al., 2017; Zhong et al., 2021) and the adaptive expectation-maximization algorithm (Tagare et al., 2010; Scheres, 2012) have been applied to image alignment. Because refinement of the entire structure has the disadvantage of averaging 2D images from various conformations, several additional refinement methods for determining the structure of flexible parts have been proposed. Multibody refinement in RELION considers complexes as a group of independent rigid bodies and calculates the sum of the rigid bodies as a flexible complex (Nakane and Scheres, 2021). The 3D variability analysis (3DVA) implemented in cryoSPARC fits a high-resolution subspace model to a flexible map (Punjani and Fleet, 2021). CryoDRGN reconstructs 3D maps into the several classes using deep neural networks (Zhong et al., 2021; Kinman et al., 2023). Murshudov et al., 2011 reviewed software for model fitting (e.g., Coot (Emsley et al., 2010), Jiggle Fit (Casanal et al., 2020), and Morphing (Casanal et al., 2020)) and refinement (e.g., ProSMART (Nicholls et al., 2014) and REFMAC (Brown et al., 2015). To evaluate the model, a FSC curve is obtained by calculating the correlation between two subsets, each including an independent half of the complete image set (Murshudov, 2016). In the FSC curve, the resolution at FSC values of 0.143 and rarely 0.5 is defined as the resolution of the model (Rosenthal and Henderson, 2003). An additional sharpening step that specifies the B-factor value improves the EM map interpretability (Fernandez et al., 2008). The local resolution of the map is processed and visualized using UCSF ChimeraX (Goddard et al., 2018; Pettersen et al., 2021).

As mentioned in Section 3.1, the use of scaffolds is an effective solution for increasing the size of biological samples and overcoming the problem of indistinguishable signals from noise in small targets. Early methods that used scaffolds in sample preparation for cryo-EM were the same as those used in crystallography: inserting a known epitope sequence into the target and forming an antibody/nanobody/Fab–target complex. However, improvements in antibody engineering techniques using computational approaches have allowed for the efficient design of target-specific antibody/nanobody/Fab and their application to the structural determination of target molecules. Recently, the use of the megabodies and Legobodies has been proposed as the most promising approach for cryo-EM studies. Additionally, ankyrin repeats are used in cryo-EM to prepare the target-specific scaffolds, called designed ankyrin repeat proteins (DARPins), by forming oligomers with scaffold cores. Scaffolds affect not only the target size but also the diversity of the orientation of the samples, and comprehensively enhance the resolution. In this section, we describe scaffolds and their applications in structural determination.

A BRIL-based scaffold was designed to reveal different states of GPCRs by various regulators using cryo-EM (Bernhard and Che, 2024). The mBRIL, the less flexible version of BRIL, was developed to reduce idiosyncratic behavior in BRIL-inserted GPCR (Guo et al., 2024). With the mBRIL, the extended cytosolic helix at C-terminus using ALFA tag (Gotzke et al., 2019) and the scaffolds between ALFA tag and mBRIL were applied to identify the complex structures of β2-adrenergic receptor with FDA-approved drugs, olodaterol and formoterol (Guo et al., 2024).

Laverty et al., 2019 developed the first megabody Mb38, the chimeric protein comprised two parts: 1) a specific nanobody to the α1-subunit of the γ-aminobutyric acid receptor subtype-A (GABA_A receptor) and 2) the scaffold protein from the extracellular adhesin domain of Helicobacter pylori HopQ. By using Mb38, the structure of the full-length human α1β3γ2L GABA_A receptor was determined at 3.2 Å resolution in the lipid bilayer. Masiulis et al. also reported the five complex structures of the full-length human α1β3γ2L GABA_A receptor with picrotoxin, bicuculline, GABA, alprazolam, and diazepam in separate lipid nanodiscs using Mb38 (57 kDa) (Figure 4A) (Masiulis et al., 2019). In 2020, using the variant of Mb38, Mb25 (56 kDa), the structure of the human β3 homopentameric GABA_A receptor was identified at 1.7 Å resolution (Nakane et al., 2020). Later, the megabodies targeting homopentameric GABA receptors were further improved to increase the bulkiness of samples and reduce the biased distribution of orientation of membrane proteins (Uchanski et al., 2021). Other types of the megabodies have been developed to target small size of membrane proteins. The human sodium/bile acid cotransporter (NTCP, 36 kDa) was structurally identified using Mb91, a fusion protein of the nanobody with E. coli glucosidase YgjK (89 kDa) at 2.88 Å resolution (Goutam et al., 2022). Mb177, similar to Mb91, was also developed and used to determine the structure of human Hedgehog acyltransferase (HHAT, 58 kDa) at 2.7 Å resolution (Coupland et al., 2021).

Another approach for solving this problem is to use Legobodies. Legobody is an assembly of three components: 1) target-specific nanobody (≈14 kDa), 2) the nanobody-binding Fab (≈49 kDa), 3) the maltose-binding protein (MBP)-based fusion protein (MBP with nanobody-binding protein A domain C (PrAC); V_H (in Fab)-binding protein A domain D (PrAD); and C_H (in Fab)-binding protein G (PrG)), called MBP–PrAC–PrAD–PrG (59 kDa) (Wu and Rapoport, 2021). The MBP–PrAC–PrAD–PrG protein reduces the flexibility of target–nanobody–Fab trimer in solution and increases its size. By binding to the heterotrimeric Legobody, the target protein obtains approximately an additional 120 kDa of its original size. Wu and Rapoport, 2021 used this system to solve the structures of both endoplasmic reticulum (ER) lumen protein-retaining receptor 2 (KDELR2, ≈30 kDa) and SARS-CoV-2 spike protein receptor-binding domain (RBD, ≈25 kDa). The Legobody was also used to determine the structure of the inner mitochondrial membrane protein uncoupling protein 1 (UCP1). UCP1 is activated by fatty acids and small molecules such as 2,4-dinitrophenol, and negatively regulated by purine nucleotides, ATP, and GTP (Kang and Chen, 2023). Using Legobodies including newly screened UCP1-specific nanobody (called sybody 12F2), Kang and Chen identified the structures of UCP1 (32 kDa) as three different states: nucleotide-free, 2,4-dinitrophenol-bound, and ATP-bound. Although the model structures of the Legobodies are not shown in the models deposited in the PDB (8HBV, 8HBW, and 8J1N), their maps are clearly described in the EM maps in EMDB (EMD-34644, EMD-34645, and EMD-35928).

In addition to antibody-like proteins, engineered ankyrin repeats have been used to design target-specific binding partners (Li et al., 2006). Initially, based on the ankyrin repeat, a target-specific DARPin was screened from the DARPin library and used as a crystal chaperone and biosensor (Pluckthun, 2015; Boersma, 2018). Due to the small size of DARPin, it should be used with scaffold cores such as aldolase in structure determination using cryo-EM (Liu et al., 2019; Yao et al., 2019). The use of the DARPin–scaffold core fusion increases the sample size by binding of the DARPin–scaffold core to the target and by forming oligomers of targets, mediated by the scaffold core. The structure of green fluorescent protein (GFP, 26 kDa) was identified using aldolase-fusion DARPin (tetramer, D2) at 5–8 Å resolution (Yao et al., 2019). Later, improvement of the DARPin–nanocage fusion scaffold (called DARP14, heterotetracosamer as a dodecameric heterodimer of DARP14 subunit A (DARPin-conjugated form, 34 kDa) and subunit B (DARPin-free form, 14 kDa)) enhanced the resolution of GFP to 3.8 Å (Liu et al., 2018a; Liu et al., 2019).

SBDD studies require the screening of several designed drugs. Hence, cost-effective and rapid methodologies are important to render the process more efficient. As shown in Table 1, to overcome the disadvantages of the cryo-EM workflow, such as lower throughput of processes compared to crystallography, methodologies have been developed for more efficient processes, such as automation of sample preparation (Koning et al., 2022), data collection steps (Tan et al., 2016) and minimization of required image numbers for structure determination (Cushing et al., 2023).

One of the recent achievements is automation of data acquisition. Smart EPU Software is available for high-throughput data acquisition with autoloader-equipped cryo-TEMs (Drulyte et al., 2022). Using the autoloader, screening and collecting images from maximum 12 grids were available. It decreases the time consumed for sample loading and environment setting such as vacuum condition and temperature. The EPU software also contains a rapid data acquisition mode that allows unattended screening of multiple grids. The “Fast Acquisition” mode in EPU utilizes beam tilt instead of stage movement to each point. Neutralizing SARS-CoV-2 antibodies specific to the RBD domain were generated, and 12 Fabs with spike proteins in complex forms were selected and subjected to data collection for structure determination using the EPU software in 48 h. After data processing, twelve sub-3 Å structures of Fabs with spike protein were reconstructed. As well as EPU, SerialEM also uses beam-tilt compensation algorithm for fast image collection (Takaba et al., 2020). EPU, SerialEM and Leginon (Cheng et al., 2021) provide condition of holes and squares of sample grids, and allow to screen and collect data in promising area of grids, and finally reduce the time for data acquisition. Ptolemy software brought the full process of data collection by automation using machine learning and computer vision algorithms (Kim et al., 2023).

Cushing et al., 2023 tested the correlation between the number of images for 3D model reconstruction and resolution to determine the structures for SBDD. The cyclin-dependent kinase (CDK)-activating kinase (CAK) heterotrimeric protein complex was formed by CDK7, MAT1, and cyclin H and affected cell division and cell growth by regulating transcription initiation (Sava et al., 2020). Owing to its important role in cellular pathways, the CAK complex is a potential target for cancer drugs and antivirals (Hutterer et al., 2015). Cushing et al., 2023 recently reported the complex structures of CAK complex with 12 different inhibitors using the rapid sample screening strategy. They initially collected ≈500 particles/grid from 26 grids using protein complexes with 12 inhibitors for 1 h and selected 13 grids among them based on processed results including refinement. The selected 13 sets were subsequently used to collecting ≈2,000 particles/grid for 4 h. This strategy suggested that a small number of particles was enough to screen and select the samples for further process. Using the fast-screening strategy, they determined a dozen of structures at 3.5–4 Å resolution daily using only a 200 kV cryo-TEM (Figure 4B).

In cryo-EM, samples are selected using the particle-picking process and used as input data. To exploit this advantage, the desired samples are selected and enriched using functionalized grids. For example, the grid was coated with affinity ligands and antibodies to capture samples (Llaguno et al., 2014; Yu et al., 2016). Functionalized grids successfully enrich the number of target proteins on the grid surface, enabling the use of low-concentration samples such as viral particles or membrane proteins, which are the main targets in SBDD studies, without the need to increase expression levels (Yu et al., 2016; Wang et al., 2020b).

For this purpose, high-affinity selective ligands, such as nickel–nitrilotriacetic acid (Ni–NTA) to His-tag, biotin to streptavidin, and antibodies to Protein A/G, have been successfully used as functional groups coated on the grid (Llaguno et al., 2014; Earl et al., 2017) (Figure 4C). Yu et al., 2016 used antibody-based affinity grids to solve the Tulane virus structure at 2.6 Å resolution using a low concentration of the virus. Using this approach, researchers detected the target sample bound to antibody–coated grids, suggesting that this approach is useful for target samples that are difficult to prepare and have low yields. Recently, Cheng et al., 2023 introduced dual-affinity graphene grids prepared with two ligands of different affinities: Ni-NTA and polyethylene glycol–biotin. By targeting different sites using dual-affinity graphene grids, they obtained 20S proteasomes. They also solved the spike protein of SARS-CoV-2. Dual-functionalized grids have higher specificity and affinity for target sample molecules than single-affinity-functionalized grids, leading to their balanced distribution on the grids. Therefore, the dual-affinity graphene grid has been claimed to be an effective approach for preparing ideal grids for structural determination of drug targets using cryo-EM.

More generally, functional groups were used to modify cryo-EM grids. Agard’s group applied amino/PEG-amino graphene oxide in their research (Wang et al., 2020b). Multifunctional graphene for cryo-EM grids, functionalized with amine, carboxyl, thiol and phenyl groups, were also developed (Naydenova et al., 2019b). Using the functionalized grids, Naydenova and colleagues successfully enhanced randomness of particles orientation, and determined structures of ribosome and apoferritin at high-resolution. In addition to non-covalent functionalization, Nickl et al., 2023 developed graphene-based grid employing covalent bonding to stabilize specimens.

SBDD studies rely on the identification of drug-binding pockets in proteins and ligands in the binding pockets (Figure 1). However, identification of ligand-binding pockets is difficult if the protein has undefined pockets (Shelke et al., 2010). To overcome this limitation, fragment-based approaches have been widely used in SBDD studies (Murray and Rees, 2009). Therefore, FBDD is effective for identifying novel ligand-binding pockets and developing potent drugs. In this approach, drug-like molecules are generated from small chemical fragments that bind to drug target proteins with low affinity (Li, 2020). Through structural studies of target proteins complexed with small compound libraries, target-bound fragments and their binding sites can be identified (Wang et al., 2023).

This method was initially developed for X-ray crystallography but was soon used in NMR-based drug screening. Cryo-EM has become a powerful tool for FBDD studies owing to technical advancements that allow the high-resolution of structures. In a recent study, Saur et al., 2020 showed that cryo-EM can be utilized for FBDD studies by identifying two different target proteins, β-galactosidase and pyruvate kinase M2 (PKM2) (Figure 4D). β-galactosidase is a homotetrameric protein with 465-kDa molecular weight that plays a role in the hydrolysis of lactose to glucose and galactose (Bartesaghi et al., 2015). The cryo-EM structure of β-galactosidase at a resolution between 2.2 and 2.3 Å was determined using three different fragment-sized ligands. The ligand structures at the binding site and the conformational changes were clearly detected in the density maps, indicating that cryo-EM has the potential to guide FBDD research. In another study, the capacity of cryo-EM for fragment screening was demonstrated using an oncology target, PKM2, which is involved in the conversion of phosphoenolpyruvate to pyruvate. PKM2 is a potential therapeutic target because of its involvement in cancer development (Zhu et al., 2021). To demonstrate the applicability of cryo-EM for fragment screening, 68 different ligands were used at high concentrations to form complexes with PKM2. The structures of these complexes were solved after 68 days of data collection, followed by 2–3 days of data acquisition depending on the sample. By preparing these structures, they indicated that a 3.2 Å resolution was adequate for determining the binding modes of fragment-sized molecules and clearly showed that the fragments were easily differentiated from noise in the density maps. After separately determining the protein-ligand structures, the structures of the protein-ligand complexes were determined using fragment cocktails consisting of four different ligands. Therefore, the use of compound cocktails for fragment screening with cryo-EM is an efficient and timesaving method for increasing throughput.

Antibodies have been used as therapeutic tools using various approaches, such as blocking the binding interface and inhibiting activity (Carter and Rajpal, 2022). In particular, antibody–drug conjugates and antibody-based proteolysis-targeting chimeras (AbTACs) are emerging fields of antibody therapy (Fu et al., 2022; Zhao et al., 2022). A critical step in antibody-based therapeutic strategies is the acquisition of appropriate antibodies with high target specificity and affinity. The most popular approach for developing new antibodies is to screen antigen-binding complementarity-determining region (CDRs) sequences using a phage library (Alfaleh et al., 2020). Other approaches include analyses of genomic sequence information from target-responding B cells extracted from infected samples and direct peptide sequence information of antibodies in the serum using mass spectrometry (Ionov and Lee, 2022). Methodologies for computer-aided antibody design have also been developed, such as antibody–antigen docking (Hummer et al., 2022). Although diverse approaches in antibody design have accelerated development speed, structural validation of the designed antibody on-target, as expected, is still a rate-limiting process using crystallography. Owing to the lack of a crystallization step, the cryo-EM approach is advantageous for determining the structures of the target–antibody complex at high resolution. Moreover, after acquiring the initial structure of the target–antibody complex, further improvements based on structural information are easily achieved. In this subsection, we describe new techniques for supporting the design of antibodies using cryo-EM: epitope mapping and the direct determination of antibodies (Nilvebrant and Rockberg, 2018; Schardt et al., 2022).

The cryo-EM-based polyclonal epitope mapping (cryo-EMPEM) method was developed in the late 2010s for screening purposes. Cryo-EMPEM determines the sequence of polyclonal antibodies from the electron density map of the antibody-target complex, unlike previous approaches such as B cell sequence analysis or mass spectrometry analysis of polyclonal antibodies. This approach integrates antibody selection and structural determination. As a recent example of cryo-EMPEM, a novel approach was described for antibody discovery in early 2022, targeting human immunodeficiency virus-1 (HIV-1) envelope glycoprotein (Figure 4E) (Antanasijevic et al., 2022a). Serum-derived antigen-bound polyclonal antibody sequences were identified using cryo-electron microscopy density maps. For this purpose, the animals were immunized with labeled antigens to produce antigen/antibody complexes that were then analyzed using cryo-EM to identify epitopes. After reconstruction, the amino acid sequences of the antibodies were determined from density maps using a novel algorithm developed by the same research group. The main aim of this algorithm is to match density maps with the antigen-binding-specific B-cell next-generation sequencing database to rapidly obtain epitope information that can be used for the rational design of therapeutics. By the end of 2022, the same research group adapted this approach to analyze polyclonal antibody responses using whole viral particles for non-enveloped viruses with icosahedral capsids, indicating that this method is an effective way to identify antibodies and map epitopes (Antanasijevic et al., 2022b).