Technological advances in super-resolution microscopy to study cellular processes

Abstract

Since its initial demonstration in 2000, far-field super-resolution light microscopy has undergone tremendous technological developments. In parallel, these developments have opened a new window into visualizing the inner life of cells at unprecedented levels of detail. Here, we review the technical details behind the most commonly used implementations of super-resolution microscopy and highlight some of the recent, promising advances in this field.

eTOC Blurb

Bond et al provide a technologically focused review on common super-resolution light microscopy approaches as well as recent, promising technical developments in this field. The review particularly highlights emerging advanced imaging approaches that hold promise for visualizing the molecular architecture of cells with high spatiotemporal resolution.

Introduction

Over the years, genomic, transcriptomic and proteomic approaches have transformed our understanding of the molecular makeup of tissues under physiological and pathological conditions at the single cell level (Stuart and Satija, 2019). However, cells are not a soup of randomly distributed proteins and nucleic acids. Proteins rarely function in isolation inside the cell but have to form complexes with other proteins at the right place and time (Marsh and Teichmann, 2015). Our genomic DNA is intricately folded to bring together distant genomic sequences in 3D space to regulate gene function (Rowley and Corces, 2018). Myriad molecular mechanisms are at play including phase separation, inter- and intra-molecular interactions to coordinate a temporally and spatially compartmentalized sub-cellular organization. Bulk –omics methods often fail to capture this dynamic spatiotemporal organization of proteins and nucleic acids that helps define a cell’s molecular state. Microscopy, in particular light microscopy, remains an indispensable tool to visualize this dynamic sub-cellular organization. However, until the last decade, far-field light microscopy remained limited in its spatial resolution to ~200–300 nm due to diffraction. To place that number in context, the synaptic cleft between two neuronal processes is ~20 nm and the nucleosome that forms the repeating unit of chromatin is only ~10 nm. The impact of super-resolution microscopy methods that overcome this limitation can therefore not be overstated. The super-resolution revolution started with the demonstration of STED (Stimulated Emission Depletion Microscopy) in 2000 (Klar et al., 2000) and continued with the subsequent development of single molecule localization microscopy (SMLM) techniques STORM (Stochastic Optical Reconstruction Microscopy) (Rust et al., 2006) and PALM/fPALM (Photoactivated Localization Microscopy and Fluorescence Photoactivation Localization Microscopy) (Betzig et al., 2006; Hess et al., 2006) in 2006. Since then, there has been an explosion of technological developments that expanded these methods to 3D imaging, multiplexed/multi-color imaging and multimodal imaging. Here, we overview super-resolution microscopy methods with a particular focus on the most commonly used and widely adopted implementations and highlight some of the exciting recent technological developments that hold high promise for studying biological processes at a level of unprecedented detail. While this review is meant to be a technologically focused review, we provide several examples of biological applications of the various methods. In addition, we discuss the pluses and minuses of each technology to guide the reader in choosing the right method. We also provide guidance in the form of a Table that lists the most common range of spatial and temporal resolutions accessible to each technology as well as other important parameters that should be considered by researchers seeking to apply these methods to their biological question of interest. Finally, we refer the readers to several recent reviews that provide a focus on biological applications of super-resolution microscopy and further guidance on how to choose the best method (Schermelleh et al., 2019; Sigal et al., 2018; Valli et al., 2021).

Super-resolution modalities

STED and RESOLFT

STED (Klar et al., 2000) was the first far-field super-resolution light microscopy modality to be demonstrated. To break the diffraction limit, the sample is scanned not only with an excitation beam, but also a co-aligned STED beam that forces excited fluorophores within a region where the STED beam has high intensity back to a non-fluorescent state through stimulated emission. When the STED beam is shaped like a donut in the lateral dimension (Klar et al., 2001) (Figure 1A(i)), only the molecules at the center of the donut, where the STED beam intensity is minimum, emit observable fluorescence, effectively shrinking the point spread function (PSF) of the illumination. As scanning coordinates are pre-determined with nano-scale precision, the fluorescence intensity at each scanning position is recorded to build a super-resolved image down to ~20 nm in resolution when using common fluorophores (Figure 1).

Figure 1: STED and RESOLFT.

(A) (i) STED depletion beam (red) geometry in lateral (x,y) dimension overlayed with the excitation beam (green), and (ii) cross-section of the z depletion beam to improve axial resolution (Hell, 2015; Willig et al., 2007). The intensity at center of the STED beam is zero, which allows for the few molecules there remain fluorescent (yellow dots) while the surrounding molecules are forced to the dark state (black dots). *temporal resolution scales inversely with the size of the field-of-view. (B) Two-color STED image of Xenopus nuclear pore complex. Scale bar, 500 nm (Göttfert et al., 2013). (C) 3D volumetric RESOLFT image of mitochondria in the axial direction in a U2OS cell expressing rsEGFP2-Omp25. Upper two panels, enhanced confocal images without 3D off-switching. Lower two panels, 3D RESOLFT images. Scale bar, 500 nm (Bodén et al., 2021). (D) Axial mitochondria dynamics (rsEGFP2-Omp25) captured by 3D RESOLFT (Bodén et al., 2021). (E) The scheme of parallelized RESOLFT (MoNaLISA) imaging the neuronal ER labeled with rsEGFP2-Sec61β (Damenti et al., 2021). The imaging setup includes parallelized foci switch-on/off and readout as illustrated. These foci only need to scan a limited fraction of the whole field (indicated by the gray lines) to cover a large 50 μm² field of view. (F) An extensive ER network (rsEGFP2-Sec61β) in a hippocampal neuron, super-resolved by parallelized RESOLFT. Scale bar, 5 μm. Inset: zoomed in dynamics of ER at 0.2 Hz in the ROI II. Scale bar, 500 nm.

The critical component in STED is to effectively suppress the fluorescence outside of the center of the illumination, which requires high STED beam intensity (MW-GW/cm²) to increase the probability of the STED beam to interact with every electron in the excited state during the lifetime of the excited state. The high intensity and repeated cycles of excitation and depletion cause considerable photobleaching, which degrades image quality, and therefore it is important to choose appropriately bright and photostable fluorophores (Zheng and Lavis, 2017) or replenishable labeling strategies (Nicolai et al., 2011; Spahn et al., 2019). Implementation of STED using continuous wave (cw) lasers (cw-STED) reduces the STED beam illumination intensity compared to the use of STED pulses, hence reducing photodamage albeit with a compromise in spatial resolution (Willig et al., 2007). More recent implementations of STED use a sample adaptive illumination strategy to minimize photobleaching. For example, in RESCue-STED, scan beams are shutdown in regions of the sample that are not occupied by fluorophores (Staudt et al., 2011). MINFIELD is another implementation that restricts the scan field to a region about the same size as the donut minimum “valley” to avoid scanning the high intensity crests over the fluorophores to reduce photobleaching (Gottfert et al., 2017). In a related approach, Dynamic Intensity Minimum (DyMIN) (Heine et al., 2017), STED beam power is gradually increased as a fluorophore is approached and then gradually decreased during the scan away from the fluorophore, once again reducing photobleaching without limiting the field of view to a small region like in MINFIELD.

The shape of the STED beam in the focus can be adjusted to allow resolution improvements in 2D as well as in 3D. To reach an isotropic improvement of resolution, IsoSTED integrates the 4Pi geometry by wedging a sample between two opposing objective lenses and illuminating the sample simultaneously from both sides (Schmidt et al., 2008). This setup requires excellent alignment of the objectives and some adaptability of the sample preparation but yields significant improvement especially in the axial resolution. Recently, adaptive optics have further improved the axial resolution of isoSTED while imaging whole cells and tissues (Hao et al., 2021a; Velasco et al., 2021). With careful choice of fluorescent spectra, dual-color imaging has also been achieved and applied in many biological settings by using one STED beam for simultaneous fluorescence depletion (Carravilla et al., 2019; Damenti et al., 2021; Eggeling et al., 2009; Göttfert et al., 2013; Große et al., 2016; Moors et al., 2021). Remarkably, by combining reversibly switchable fluorescent proteins with organic fluorescent dyes, three-color imaging by STED can be achieved (Willig et al., 2021). Numerous investigators have successfully used STED in fixed and live samples across a wide range of biological contexts, such as observing neuronal actin dynamics or 3D imaging of extracellular space in live mouse brain, mapping of large-scale protein targets in dendritic spines, tracking cellular organelle morphological changes, characterizing lipid molecule movement on the membrane, and evaluating the accessibility of HIV-neutralizing antibodies to single virions. (Berning et al., 2012; Bottanelli et al., 2016; Carravilla et al., 2019; Eggeling et al., 2009; Große et al., 2016; Helm et al., 2021; Schroeder et al., 2018; Tønnesen et al., 2018). However despite recent strategies that reduce photobleaching, photobleaching/phototoxicity remains a concern for the most common implementations of STED microscopy. In addition, in the absence of specialized implementations of STED that use parallelized illumination (Bingen et al., 2011), the acquisition speed of STED is low when imaging large fields of view.

RESOLFT (reversible saturable optical fluorescence transitions) (Hofmann et al., 2005) is another method that uses a similar principle as STED but uses photo-switchable proteins that can change conformation upon exposure to different wavelengths of light to achieve on- and off-switching, instead of stimulated emission. Using the same donut-shape design, the off-switching requires much lower intensity (W-kW/cm²) (Figure 1), which supports continued, gentle imaging in a wide range of biological settings including mammalian cells and Drosophila tissue to give a few examples (Grotjohann et al., 2012; Schnorrenberg et al., 2016; Tiwari et al., 2015). The tradeoff for the gentle laser power of RESOFT is that it takes longer (millisecond pixel dwell time, Table 1) for fluorescent proteins to switch off whereas the depletion by STED is instantaneous (~200 psec, see Table 1 for comparative properties of various super-resolution modalities that should be considered during experimental design).

Table 1:

Comparison of super-resolution techniques and parameters that should be considered in choosing a super-resolution modality.

Technique	Commercial Availability	Equipment Cost	Lateral Spatial Resolution	Axial Spatial Resolution	Temporal Resolution	Sample penetration depth in standard setup	Specialized fluorophores	Difficulty of Data Analysis
STORM/PALM	High	Medium	~20–50 nma	~40–100 nmb	Seconds-Minutes	Several μm	Yes	High
DNA-PAINT	Medium	Medium	~20–50 nma	~40–100 nmb	Minutes-Hoursc	Several μm	No	High
ROSE/SIMFLUX	Low	Medium/High	~5–20 nma	N/A (implemented in 2D)	Minutes-Hours	Several μm	Nod	High
ROSE-Z	Low	Medium/High	~2–5 nma	~2–5 nm	Minutes-Hours	Several μm	Nod	High
ModLoc	Low	Medium/High	~3–5 nma	~7–10 nm	Minutes-Hours	Several μm	Nod	High
Expansion Microscopy	High	Low-Mediume	~25–70 nmf	~25–70 nm	Seconds-Minutes	Variableg	No	Low
STED	High	High	~20–50 nm	~40–100 nmh	10 ms-Minutesi	Typically 10–15 μm, up to 120 μmj	No	Low
RESOLFT	Medium	High	~20–50 nm	~40–100 nm	Seconds-Minutesi	Several μm	Yes	Low
Parallelized RESOLFT	Low	High	~45–65 nmk	~50–80 nmk	Seconds-Minutes	Several μm	Yes	Low-Medium
MINFLUX	Medium	High	~1–3 nm	~1–3 nm	Down to 80 μsecondsi	Several μm	Yes	High
SIM	High	Medium	~100–150 nm	~250–350 nm	Sub-seconds	Several μml	No	Medium
TIRF-SIM	High	Medium	~85–120 nm	~100–200 nm	Sub-seconds	200 nm	No	Medium
ISIM	High	Medium	~140–200 nm	~300–400 nm	Down to 10 ms	20 μm	No	Medium
Airyscan	High	Medium	~120–180 nm	~350–450 nm	Milliseconds-Secondsi	Several μm	No	Low
SOFI	High	Low	~75–175 nmm	~200–300 nmm	Sub-seconds-Seconds	Several μm	Yes	High
SRRF	High	Low	~60–150 nm	Diffractionlimited (no improvement)	Seconds	Several μm	no	high

^a.Values depend on many parameters, importantly the localization precision, which is dependent on the number of photons collected and the fluorophores used. Ranges correspond to most commonly used and bright fluorophores.

^b.Values correspond to the most commonly used PSF engineering implementations (e.g. cylindrical lens) and can further be improved using interferometric approaches (iPALM, 4Pi-SMS and W-4Pi-SMSN)

^c.Values are dependent on number of targets to image; with improved buffers and DNA-sequences this can be decreased to seconds-minutes

^d.These methods can be implemented with STORM fluorophores or DNA-PAINT approaches. When implemented with DNA-PAINT, they do not require specialized fluorophores.

^e.ExM can be done on a simple widefield microscope, but generally a confocal microscope is used due to sample size and thickness

^f.70 nm refers to the most common ~4.5X sample expansion. With improved gel recipes or iterative expansions, larger expansions are possible, bringing resolution down to 25 nm

^g.ExM imaging generally requires many Z-stacks based on increased sample size

^h.Values correspond to most common implementation of 3D STED and can further be improved with the use of 4Pi geometry (isoSTED)

^i.Temporal resolution inversely scales with the size of field of view.

^j.Deep sample penetrance has been achieved with high N.A glycerol immersion lenses (Urban et al., 2011)

^k.Lateral and axial resolution based on numbers reported in (Masullo et al., 2018) and (Bodén et al., 2021), respectively.

^l.Up to 80 um depth has been demonstrated with adaptive optics (Lin et al., 2021).

^m.The resolution here is based on 2^nd to 4^th order statistical analysis, the most common range for SOFI.

A series of developments to make fluorescent proteins brighter, faster-switching, more photostable, and/or decoupling excitation from photo-switching has made RESOLFT more desirable for live-imaging (Jensen et al., 2020). Imaging throughput has further been improved by using massively parallelized illumination (Bodén et al., 2021; Chmyrov et al., 2013) as well as similar sample adaptive illumination strategies as STED that enable only scanning regions of the sample that are occupied by fluorophores (Dreier et al., 2019). In addition, MoNaLISA (Molecular Nanoscale Live Imaging with Sectioning Ability) used different On- and Off-switching light patterns with optimized shape and periodicity to effectively balance imaging speed, field-of-view and 3D resolution (50 × 50 μm² at 0.2–1.3 Hz with 45–65 nm resolution) (Masullo et al., 2018) (Figure 1). This setup allowed the investigators to track the ER and mitochondria contact dynamics in different compartments of the neuron simultaneously at seconds and minutes time resolution (Damenti et al., 2021). These recent iterations of RESOLFT that use massively parallel, sample adaptive and optimized light illumination patterns, although not commercialized, hold great promise for live-imaging applications of super-resolution microscopy as they achieve imaging of large fields of view at high spatiotemporal resolution with reduced photobleaching and phototoxicity.

SMLM

SMLM utilizes the principle that a single fluorescent molecule can be localized by pinpointing its center of emission, resulting in a point estimate with a precision on the order of several nanometers. However, this approach requires that ideally only a single fluorophore is imaged in a given diffraction limited volume at a time. SMLM methods overcome this challenge by detecting a “blinking” fluorescent signal (see below). These localizations are then used to reconstruct a higher resolution image after sufficient rounds of blinking have allowed for all labels on the target structure to be localized. This generally necessitates several thousands of frames to be acquired for successful reconstruction of the final image. Different SMLM modalities achieve the required sparseness of the single molecules using different mechanisms of “blinking” as discussed below:

STORM/PALM

STORM and PALM utilize similar principles, and historically the major difference between them has been the type of fluorophore used. Initially, PALM utilized genetically encoded photoactivatable fluorescent proteins (Betzig et al., 2006), while STORM used organic fluorescent dyes (Rust et al., 2006). In both cases, the photophysics of the fluorophores allows for switching between two states (on/off or two different colors) for precise localization of a small number of blinking events at a time, resulting in a resolution reaching ~20 nm in x and y. Both methods have since been extended to 3D imaging with additionally improved z-resolution. One class of 3D methods use various ways of point spread function (PSF) engineering, in which the shape of the PSF is dependent on the z-position of the fluorophore (Hajj et al., 2014; Huang et al., 2008) Alternatively, 3D near-isotropic resolution can be achieved via the use of a dual-objective geometry combined with PSF engineering (Xu et al., 2012) or interferometric methods including iPALM (Shtengel et al., 2009), 4Pi-SMS (Aquino et al., 2011), and W-4PiSMSN with extended depth penetration over a whole cell (Huang et al., 2016).

STORM and PALM have been instrumental for many biological discoveries for over a decade ranging from the discovery of the Membrane-Associated Periodic Skeleton (MPS) in neuronal axons (Xu et al., 2013), the visualization of the folding patterns of DNA in the nucleus (Bintu et al., 2018; Neguembor et al., 2021; Ricci et al., 2015), the determination of the functional clustering of membrane receptors (Rossy et al., 2013) to the characterization of protein aggregates (Gyparaki et al., 2021). For a detailed review on the technical details behind STORM and PALM and their biological applications the reader is directed to a recent primer on this topic (Lelek et al., 2021). STORM and PALM are compatible with traditional widefield microscopes, making them accessible to many researchers, but they are not without limitations. With a few exceptions, imaging is generally limited to 2–3 colors in thin, fixed samples. These techniques have been combined with emerging advanced imaging methods to overcome some of these challenges (see section on Combined and Correlative Approaches).

DNA-PAINT (DNA points accumulation for imaging in nanoscale topography)

DNA-PAINT relies on the same concept of blinking as STORM and PALM, but the method for achieving blinking is unique. In DNA-PAINT, the target is labeled with a short DNA oligonucleotide (the docking strand) rather than a fluorophore (Jungmann et al., 2014). The fluorophore is floating freely in solution, tethered to the complementary, short DNA oligonucleotide (the imager strand). When the imager strand transiently binds to the docking strand and becomes temporarily immobilized, this event is registered as a blink. A major advantage is that the imagers can be washed out after a round of imaging, and hence the number of targets that can be imaged is only limited by the number of unique docking strands. Recently, 124-color DNA-PAINT multiplexing has been shown via a unique process referred to as kinetic barcoding (Wade et al., 2019). This strategy separates targets based on the tightly engineered binding kinetics of the imager oligo leading to distinct “blinking” rates, pushing the limits of super-resolution multiplexing and offering a major advantage over STORM/PALM. Similar to STORM, the DNA-PAINT docking oligos can be conjugated to primary or secondary antibodies, nanobodies as well as Halo-tags (Schlichthaerle et al., 2019), making DNA-PAINT a versatile approach.

While the multiplexing capability of DNA-PAINT is a significant advantage over STORM/PALM, this method has not been without its relative downsides. One downside is the fact that docking and imager oligos can have off-target binding to genomic DNA, leading to high background signal in the nucleus and making it more difficult to image nuclear targets. To overcome this challenge, a method called L-DNA-PAINT has been developed that takes advantage of left-handed DNA (L-DNA) oligomers that do not hybridize with right handed DNA (R-DNA) (Geertsema et al., 2021). Another major downside of traditional DNA-PAINT is imaging speed, which is generally 10–15 times slower than STORM for acquisitions of the same number of frames (Table 1). This is due to the need for long exposure times to blur out the fluorescence from imagers in solution, and to avoid localizing freely floating imagers that transiently move into the field of view. Another limitation that impacts imaging speed is the diffusion limited time that it takes for an imager strand to bind to its target strand. This waiting time is inversely proportional to the imager oligo concentration, however, the imager oligo concentration has to be kept low to avoid excessive background fluorescence. Recently, to increase the number of imager binding events, improved DNA sequence design and buffer conditions were introduced, allowing for imaging times comparable to STORM (Schueder et al., 2019; Strauss and Jungmann, 2020). However, these improvements do not yet address the problem of transient imager diffusion into the field of view without stable binding. This problem can be overcome with the use of fluorogenic dyes, which only fluoresce when bound to their target sequence (Chung et al., 2020). These can dramatically reduce background and off-target signal, enabling faster acquisition times. Combined, these advancements show promise for DNA-PAINT to offer super-resolution imaging of many targets with comparable imaging speeds to other SMLM techniques. A final downside is that DNA-PAINT is not generally compatible with live cell imaging due to both the long acquisition times of traditional DNA-PAINT and also the need for the imager oligo to cross the plasma membrane to visualize intracellular targets.

As DNA-PAINT technology continues to mature, we look forward to improved commercial availability of DNA-conjugated probes, including primary antibodies and Halo-tags, and widespread implementation of fluorogenic probes and oligo sequences to improve speed. We believe these advancements will lead to further adoption of DNA-PAINT as a routinely used and robust SMLM modality. As the method gains popularity it will be exciting to watch further advancements in both the technique and in the biological insights gained from its use.

Image analysis and quantification in SMLM

The tools for analyzing traditional optical microscopy images that consist of pixel-based intensity information have been developed over many years and continue to evolve. Several open-source software packages are available for analyzing these types of images, including ImageJ (Schneider et al., 2012), CellProfiler (Carpenter et al., 2006) and others and these tools are readily applicable to super-resolution images that are intensity based such as STED images. Methods like STED and RESOLFT also require minimal or no pre-processing to obtain a final image whereas SMLM requires specialized pre-processing steps. In addition, SMLM generates pointillist images consisting of a list of fluorophore positions. Therefore, post-processing of SMLM images requires specialized tools that are capable of handling this pointillist information. Consequently, image quantification in SMLM is a challenging bottleneck and large efforts have been devoted to the development of tools that are uniquely capable of analyzing SMLM data.

Pre-processing for SMLM

The first pre-processing task is to identify peaks (i.e. individual fluorophore images or PSFs) and determine their location through the estimation of the PSF’s center position (localization) (Figure 2). The localization step can be achieved through several methods including nonlinear least squares fitting to a Gaussian PSF model or maximum likelihood estimation (MLE) using a Gaussian PSF and Poisson noise model (Smith et al., 2010). However, analytical functions such as a Gaussian may not accurately capture the PSF shape. More recently, other estimation methods have been used, such as fitting to cubic splines (Babcock and Zhuang, 2017) or to experimental PSFs (Li et al., 2018). While the most precise way to localize fluorophores is to obtain sparse fluorophore images, acquisition of high-density fluorophore images can speed up imaging time. Hence, methods have also been developed to handle data containing high emitter density, including multi-emitter fitting algorithms (Holden et al., 2011), compressed sensing (Zhu et al., 2012) and more recently machine learning (Nehme et al., 2018; Speiser et al., 2021) (see below). Several open and closed source SMLM software packages are available for achieving the peak detection and localization steps. Users often face the dilemma of picking the “right” image processing platform. A recent paper evaluated the performance of several open-source software options to assess detection rate, accuracy, software usability and speed (Sage et al., 2019). This comparison serves as a resource to evaluate the tradeoff between SMLM packages in detecting and localizing fluorophores under different experimental conditions. Some of these software packages also include additional pre-processing steps such as linking the same fluorophore across multiple frames, filtering imprecise localizations and drift correction. The reader is directed to a recent primer on SMLM for more in depth reading of SMLM pre-processing (Lelek et al., 2021).

Figure 2: Single Molecule Localization Microscopy and analysis.

(A) The pre-processing tasks of SMLM data consist of: (i) peak detection and localization, (ii) linking of localizations, (iii) filtering, (iv) drift correction, and (v) rendering. Currently, there are more than 25 open-source software packages that can perform all or some steps of these tasks. Figure was adapted from (Brede and Lakadamyali, 2012) (B) Post-processing of SMLM images include: (i) clustering, (ii) segmentation, (iii) co-localization, (iv) classification, and (v) validation. (1) SR-Tesseler, developed by (Levet et al., 2015), segments SMLM images consisting of a wide-range of shapes and sizes using Voronoi Tessellation (1a), such as entire neuronal processes (1b, c) or nano-clusters (1d). (2) An up-and-coming category of SMLM analysis platforms are those that apply machine- and deep-learning algorithms. For example: ANNA-PALM, developed by (Ouyang et al., 2018), reconstructs super-resolution images (2c) from sparse, rapidly acquired SMLM data (2a,b). (3) Simulation of realistic SMLM data can be used to validate and compare the performance of available SMLM analysis platforms. For example, SuReSim, developed by (Venkataramani et al., 2016), is a software that can simulate biological structures under different experimental conditions; 3a shows the ground truth model, 3b shows the simulated image, and 3c shows the real-experimental data for a pair of crossed-microtubules labeled with antibodies.

Post-processing for SMLM

A striking result of super-resolution studies, particularly using SMLM, has been that many proteins assemble into heterogeneous nano-clusters. Thus, several methods have been developed that can segment and quantify the nature of these nano-clusters (Figure 2).

Initial methods for quantifying clusters in SMLM images took advantage of statistical tools such as Ripley’s K function or Pair Correlation Analysis (PCA) (Khater et al., 2020). These statistical tools work on the principle that randomly distributed points in space generate a known number of localizations within a certain distance of another localization. Deviations from this random distribution can thus be determined statistically. These methods are applicable to 2D and 3D SMLM data as well as multi-color SMLM data for determining whether the distribution of two proteins is correlated in space and have been successfully applied in many biological settings including determination of membrane receptor clustering (Sengupta et al., 2011). However, they only provide statistical averages and hence heterogeneities and properties of individual clusters cannot be obtained. To overcome this challenge, several cluster segmentation methods, often relying on mathematical approaches that have been in use in other fields, have been adapted for use in SMLM, including DBSCAN (Density Based Spatial Clustering of Applications with Noise) (Khater et al., 2020), Voronoi tessellation (Andronov et al., 2016; Levet et al., 2015) (Figure 2) and Bayesian analysis (Rubin-Delanchy et al., 2015). Some of these approaches, in particular Voronoi tessellation, have also been extended to determining co-localization in multi-color SMLM images (Levet et al., 2019).

While each method has its advantages and drawbacks, none of these approaches alone can inform whether the identified clusters are true biological clusters or clustering artifacts. Pseudoclusters are the result of repetitive and stochastic photoblinking of the same fluorophore and the fact that labeling stoichiometry is typically not one to one (Durisic et al., 2014a). As a result, in SMLM it is not possible to assign each localization to a unique protein, leading to clustering artifacts and over- or under-counting of protein copy number. Several approaches have been developed to overcome this problem. A recent example is DDC (Distance Distribution Correction), which relies on obtaining the true pairwise distance distribution of different fluorophores from the imaging sequence at different frame differences (Bohrer et al., 2021). However, methods like DDC that only correct for repeated fluorophore blinking cannot account for repeated localizations arising from amplifications in labeling, for example due to the use of primary and secondary antibodies. A frequently used solution to account for both problems and to determine the underlying molecular stoichiometry incorporates the use of calibration samples (Durisic et al., 2014b; Zanacchi et al., 2017) such as the nuclear pore complexes (NPCs), which have a well-defined subunit stoichiometry (Thevathasan et al., 2019). Another, easily accessible approach for accounting for clustering artifacts involves varying the labeling density, which leads to characteristic changes in the localization maps of each acquisition (Baumgart et al., 2016). DNA-PAINT offers unique advantages over STORM/PALM for quantitative analysis. While the blinking of the fluorophores in STORM/PALM is highly stochastic, the equilibrium dissociation constant (K_d) between the docking and imager strands is precise (Jungmann et al., 2016). At a given concentration of imager oligo and secondary antibody, the number of blinking events visualized per area, thus, scales linearly with the number of labels. Thus, with careful antibody titration controls, quantitative PAINT (qPAINT) (Jungmann et al., 2016) can measure the copy number of proteins more precisely. The precision of this technique can be further improved with the use of DNA-tethered nanobodies, which reliably only have one docking strand, whereas traditional secondary antibodies generally have several.

An exciting, up and coming approach to quantification of super-resolution data is the development of tools that incorporate machine and deep learning. ASAP (Automated Structure Analysis Program) is a machine learning approach that enables rapid and automated detection, segmentation, and classification based on the structural features of SMLM clusters (Danial and Garcia-Saez, 2019). Another recently employed approach is the use of Iterative Hierarchical Clustering, which is an unsupervised classification algorithm (Gyparaki et al., 2021). Machine and deep learning have also been applied to generate high-quality super-resolution images from suboptimal data. One example is ANNA-PALM (Artificial Neuronal Network Accelerated-PALM), which can reconstruct super-resolution images from sparse, rapidly acquired SMLM data and low-resolution images captured with conventional microscopes (Ouyang et al., 2018) (Figure 2). Deep-STORM (Nehme et al., 2018) and DECODE (Deep Context Dependent) (Speiser et al., 2021) are computational tools that employ a trainable deep-convolution network to generate high-resolution images in the presence of challenging signal-to-noise ratio and high emitter density. Finally, data-driven methods have been developed that are capable of converting images acquired from low-resolution systems (widefield and confocal fluorescence microscopes) into super-resolution quality images (Wang et al., 2019). While the advantages of these approaches seem attractive, whether the resulting images are artifact-free is unclear. Hence, tools to validate super-resolution images hold great importance. TestSTORM is a simulator that can be used to optimize critical imaging parameters and understand where imaging artifacts come from (Novák et al., 2017). SuReSim is another simulation software that simulates data sets of arbitrary structures allowing users to test parameters that can potentially affect the imaging outcome (Venkataramani et al., 2016) (Figure 2).

While the caveat of deep learning is the need for robust and large training data sets using synthetic or experimental data, the incorporation of these novel methods holds much promise in pushing forward the toolset available for pre- and post-processing of super-resolution data.

Intermediate resolution modalities

As the SMLM- and STED-based super-resolution microscopy techniques continue to develop towards higher resolution, other alternative intermediate resolution microscopy methods have come into light for their advantage of better temporal resolution and less phototoxicity albeit with a somewhat modest improvement in spatial resolution (typically 1.7 to 3-fold) (Table 1).

SIM (Structured Illumination Microscopy)

The most widely used SIM, or linear SIM, illuminates fluorescent samples with a fine strip pattern so the images appear with a sinusoidal brightness (Gustafsson, 2000; Heintzmann and Cremer, 1999). These images contain information of higher spatial frequencies that are otherwise not captured by the objective lens. Typically, for 2D SIM, collecting 3–5 SIM images that are phase-shifted in illumination pattern at each of the three angles that are rotated laterally is sufficient to gain resolution. In other words, 9–15 raw images are needed to reconstruct an image with enhanced resolution. The higher frequency information that represents the fine details are untangled by a deconvolution step in the Fourier space. The final image is produced through a reverse Fourier transform step, reaching an improvement in resolution by up to two-fold. SIM has gained popularity to study a wide range of biological questions. For example, a recent study found that fission position on a mitochondrion is a strong predictor of mitochondria fate, and the researchers were able to discern the fine morphological and biochemical changes before and after fission via SIM in live cells (Kleele et al., 2021). SIM has also provided new insights into the 3D organization of chromatin and the formation of topologically associating domains (TADs) within the nucleus of single cells (Szabo et al., 2018). Due to its reliance on conventional fluorophores and fluorescent proteins, SIM can be routinely used with multicolor and has also been expanded to 3D imaging (Fiolka et al., 2012; Gustafsson et al., 2008). The signal to noise ratio can be further enhanced by combining SIM with TIRF (total internal reflection) (TIRF-SIM), although the imaging depth is limited to within 200 nm near basal cell membrane (Li et al., 2015). On the other hand, grazing-incidence SIM (GI-SIM), which essentially illuminates the sample with a 1 μm light sheet parallel to the imaging plane gives an excellent penetration for visualizing organelle dynamics at millisecond time scale (Guo et al., 2018b).

Experienced developers have been continuously expanding SIM’s capabilities on both spatial and temporal resolution. For example, non-linear SIM reaches resolution down to 45 nm on biological samples (Gustafsson, 2005; Li et al., 2015; Rego et al., 2012), fastSIM, DMD-SIM and HIT-SIM reach a speed of up to 50–60 Hz (Lu-Walther et al., 2015; Sandmeyer et al., 2021; Van den Eynde et al., 2021), and GI-SIM reaches a remarkable 266 Hz frame rate (Guo et al., 2018b). SIM has also been combined with other modalities, such as lattice light sheet (Chen et al., 2014), two-photon microscopy (Ingaramo et al., 2014; Winter et al., 2014) or adaptive optics for deep sample penetration (Lin et al., 2021). We refer the readers to recent in-depth review articles on SIM for further reading (Prakash et al., 2021; Wu and Shroff, 2018).

Another commercially available alternative, Instant SIM (ISIM), offers favorable temporal resolution by employing rapid, parallelized multifoci scanning (York et al., 2013). Instead of applying a stripe pattern at multiple angles, it combines microlenses and pinhole arrays that locally scale down each excitation and emission foci, and uses optical hardware to scan and sum the raw images for an instantaneous improvement of spatial resolution (~1.4 fold). Typical imaging speed in ISIM can be up to 100 Hz, a two-fold resolution gain can be recovered after a further deconvolution step and this method has been combined with illumination geometries such as TIRF (instant TIRF-SIM) for high contrast imaging near the coverslip (Guo et al., 2018a).

One important consideration when using SIM is that artifacts can arise post-reconstruction due to various experimental and instrumental factors. Some artifacts can be visually spotted as they carry characteristic patterns in the resolved images, which can be reduced or eliminated by adjusting different parameters during reconstruction, improving signal to noise ratio in the biological sample and/or tuning/calibration of the instrument (Demmerle et al., 2017). There are accessible open source algorithms, for example, SIMcheck, an ImageJ plugin suite (Ball et al., 2015), and other advanced methods (Huang et al., 2018; Smith et al., 2021), which are useful for detecting and eliminating artifacts.

Airyscan

Airyscan is a confocal scanning microscopy-based method, and the resolution improvement (~1.7 fold) is largely through improvements in the signal to noise ratio (Huff, 2015; Huff et al., 2017). The signal is recorded by an array of 32 small point detectors (each detector element behaving as a 0.2 airy unit pinhole) that collectively maintain the light collection efficiency of a 1.25 airy unit pinhole, improving the signal to noise without compromising resolution. Through a linear deconvolution step of each of the 32 individual elements to further improve their lateral and axial resolution followed by pixel reassignment of each element back to the center position, Airyscan produces images with enhanced lateral and axial resolution (Table 1). Because Airyscan is commercially available, gentle, easy to operate (see detailed protocol by (Wu and Hammer, 2021)), compatible with multiplexing and capable of acquiring images at millisecond temporal resolution with minimal processing, it has been routinely used in a wide range of biological applications, for example, visualizing the buckling of detyrosinated microtubules in contracting heart muscle cells (Robison et al., 2016), determining how defects in cardiac cytoskeletal organization affect failing cardiomyocytes (Chen et al., 2020), or how actin cables assemble and move at mitochondria during mitosis (Moore et al., 2021).

SOFI (Super-resolution Optical Fluctuation Imaging) and SRRF (Super-resolution Radial Fluctuation)

SOFI, similar to SMLM methods, takes advantage of the blinking nature of fluorescent probes. However, unlike SMLM methods, each image in a SOFI experiment does not require a sparse subset of active fluorophores, which results in a higher emitter density, and thus faster acquisition times. Indeed, stochastically blinking fluorophores produce fluctuations in the fluorescence intensity that are detectable by carrying out higher-order statistical analysis of these temporal fluctuations in the image sequence. This will result in a spatial resolution beyond diffraction (Dedecker et al., 2012; Dertinger et al., 2009). The faster acquisition times make the method compatible with both fixed and live samples in a multicolor setting in 3D. SOFI has shown ~2–4-fold improvement in spatial resolution on structures such as subdomains on the plasma membrane, microtubules, actin, and clathrin-coated pits (Dedecker et al., 2012; Zhang et al., 2015). SOFI can also be used for biosensing experiments. For example, a new class of biosensors termed FLINC (Fluorescence fluctuation increase by contact), in which the fluorescence fluctuation behavior of the readout protein changes when in proximity to a second fluorescent protein, have been combined with SOFI to visualize dynamic biochemical activity within cells with super-resolution (Mo et al., 2017).

The SOFI image, however, does not linearly correspond to the fluorescent intensity present within each pixel. Therefore, one should be careful when interpreting fluorophore density information. Various factors can degrade the SOFI image quality due to its assumption of independent blinking, such as bleaching and laser instabilities. Artifacts can be evaluated and addressed using open access tools (Hugelier et al., 2021; Moeyaert et al., 2020; Peeters et al., 2017; Vandenberg et al., 2016).

SRRF is an analytical framework that can improve the resolution of images acquired by conventional microscopes, by calculating the radiality of the signals across the entire frame. SRRF is applicable to a wide range of experimental datasets including those acquired for SMLM or confocal microscope images and the resolution improvement depends on the nature of the dataset (Gustafsson et al., 2016). For example, microtubules can be super-resolved in live cells with ~60–120 nm spatial resolution at 1 second temporal resolution using SRRF (Gustafsson et al., 2016).

Both SOFI and SRRF can produce improved resolution in living samples without the need for lengthy acquisition times typically required in SMLM. They provide extra tools when advanced super-resolution microscopy equipment is not accessible, and can be complementary to existing super-resolution methods that have lower temporal resolution (Deschout et al., 2016). The associated analysis algorithms are open source, including parameter guidance and evaluation (Cevoli et al., 2021; Culley et al., 2018; Moeyaert et al., 2020). However, caution is needed in both SOFI and SRRF to avoid artifacts and interpret the resolved images and algorithms have been developed to address some of the pitfalls of these high-density intensity fluctuation methods (Marsh et al., 2018).

Emerging super-resolution modalities

MINFLUX and MINSTED

Despite much advancement in super-resolution, the spatial resolution of most methods does not reach true molecular scale and the temporal resolution has been a limiting factor. MINFLUX (minimal emission fluxes) (Balzarotti et al., 2017), offers a leap forward towards improved spatial and temporal resolution (Figure 3 A–B, Table 1). In MINFLUX, the fluorescent molecule’s position can be precisely determined by exposing the molecule to multiple offsetted excitation donuts. The center of the donut-shaped excitation beam has minimum intensity, but a molecule that is proximal to the center of the donut will emit fluorescence to different degrees depending on its distance from the donut minimum. The number of photons emitted at various coordinates is used to calculate the emitter position.

Figure 3: Emerging super-resolution modalities.

(A) Graphic representation of a fluorescent molecule (star) excited by a moving, donut-shaped excitation beam (green) at 4 different positions (r₀, r₁, r₂, r₃) based on the donut minimum in MINFLUX (Balzarotti et al., 2017). Each position is color-coded, and the corresponding photon counts are shown below each panel. (B) Color-coded trajectories of 30S ribosome subunit fused to mEos2 rapidly moving in a E.coli cell (transmission micrograph), captured by MINFLUX at 125 μs temporal resolution and 48 nm localization precision, by detecting on average 9 photons per localization (Balzarotti et al., 2017). Scale bar, 500 nm. (C) Working principle of ROSE-Z. The excitation laser was phase- and amplitude-modulated, split into two paths and interfered at the sample to create an interference fringe pattern. The intensity and phase of the interference fringes are modulated to switch among three different illumination patterns during imaging. In the imaging path, the fluorescence signal is recorded at different areas of the charge-coupled device (CCD) and synchronized with the illumination pattern, resulting in three subimages within a single exposure. DM, dichroic mirror. From (Gu et al., 2021). (D) Illumination patterns along Z direction and corresponding sub-images. (E) ROSE-Z imaging of Tom20 labeled mitochondria and α-tubulin labeled microtubules in a Cos7 cell. Scale bars, 5 μm. (F) General workflow for expansion microscopy. Adapted from (Wassie et al., 2019). (G) Click-ExM image showing mitochondria in a Cos-7 cell labeled with anti-Tom20, Mito-GFP and Alkyne-Choline to label choline-containing lipids. Scale bars, 10 μm in B, 2 μm in inset of B. Adapted from (Sun et al., 2021).

The capability of MINFLUX to capture highly dynamic events was first demonstrated by tracking 30S ribosomal subunit utilizing only 10 photons per localization in live E.coli cells at a localization precision of 48 nm and at time resolution of 125 μs (Balzarotti et al., 2017). To achieve an outstanding spatiotemporal resolution, additional hardware measures such as fast estimate of position were implemented to keep the fluorescent molecules near the donut center (Eilers et al., 2018). Spatial resolution can be further improved to near molecular scale by multiple iterations of scanning and estimation of the emitter location. Additionally, MINFLUX can be implemented in 3D with a precision of a few nanometers on a fixed sample (Gwosch et al., 2020; Pape et al., 2020; Schmidt et al., 2021). Due to the low photon requirement for precise localization, MINFLUX is very powerful when applied to single molecule tracking especially when the signal is dim (Schmidt et al., 2021). MINFLUX microscopes have just begun to be commercialized very recently and may become more widely applied in biology in the future. It is yet unclear if MINFLUX can retain temporal resolution with a large field of view to carry out molecular scale super-resolution imaging at high throughput, which ultimately comes back to the tradeoff between spatial and temporal resolution, a persistent challenge in microscopy.

Using a related concept to MINFLUX, MINSTED employs a donut-shaped STED beam co-aligned with the excitation beam to scan around the fluorophore to localize it (Weber et al., 2021). In contrast to MINFLUX, a molecule at the donut minimum in MINSTED emits the highest number of photons. However, MINSTED also uses the information of photons emitted when the molecule is away from the donut center during the scan, where STED beam intensity is intermediate, and this informs the subsequent scanning iterations to close in on the molecule’s position (providing down to 1~3 nm precision). One advantage of MINSTED over MINFLUX is that it provides better background suppression (Weber et al., 2021).

SIMFLUX, ROSE, and ModLoc

While MINFLUX and MINFIELD were a conceptual leap forward in bringing the resolution even closer to the scale of individual molecules, the scanning nature of these approaches constrains imaging to a small field of view. Several recent techniques have attempted to address this limitation using extended illumination for larger fields of view, while also enhancing resolution. SIMFLUX (Cnossen et al., 2020), SIMPLE (structured illumination microscopy based point localization estimator) (Reymond et al., 2019) and ROSE (repetitive optical selective exposure) (Gu et al., 2019) utilize similar principles in which the sample is illuminated with a shifting pattern of interference fringes. These fringes are shifted three times each in x and y, creating six sub-images of fluorophores with variations in fluorophore intensity. The fluorophore’s position can then be estimated by combing the relative position with respect to the shifting illumination pattern and the estimated center of the fluorescent spot. These methods achieve ~2-fold improvement in localization precision over traditional SMLM techniques in the lateral direction. Related methods like ModLoc (Jouchet et al., 2021) and ROSE-Z (Gu et al., 2021) (Figure 3C–E) also improve the axial resolution. ROSE-Z (Gu et al., 2021) implements interference fringe excitation in the axial direction in addition to inducing an astigmatism via insertion of a cylindrical lens into the imaging path. The result is an axial localization precision of roughly 2 nm, dramatically improved compared to traditional SMLM setups that utilize PSF engineering alone. In a similar approach, ModLoc (Jouchet et al., 2021) uses an illumination pattern composed of tilted fringes and the axial localization is obtained by a combined analysis of the phase of the modulated fluorescent signal produced by the shifting structured illumination and of the lateral localization obtained by centroid fitting of the PSF leading to a ~7 nm axial and ~3 nm lateral localization precision over several microns z-depth. These approaches all utilize a single objective as opposed to the previous interferometric approaches that relied on dual-objective 4Pi geometry (iPALM, 4Pi-SMS, W4PiSMSN) and are hence in principle easier to align and maintain. These relatively new techniques, while requiring specialized setups that are not yet widely or commercially available, hold promise for enhancing resolution in future biological applications.

Expansion Microscopy (ExM)

ExM relies on isotropic expansion of samples by embedding them within a swellable hydrogel, physically increasing the distance between targets to improve imaging resolution. Initially, samples were expanded in all directions by a factor of ~4.5, offering a lateral resolution of about 70 nm (Chen et al., 2015) (Figure 3 F–G). Since then, numerous improvements have increased the method’s accessibility and resolution, broadening its applicability.

In 2016, several groups introduced new protocols for ExM that could be carried out using only commercially available reagents and making ExM a highly cost-effective technique for achieving super-resolution imaging (Chozinski et al., 2016; Ku et al., 2016; Tillberg et al., 2016). Compatible with traditional wide-field or confocal microscopes and conventional fluorophores, ExM is highly accessible compared to other techniques that require more specialized setups, fluorophores, reagents and image processing tools. Recently, for example, ExM was used to investigate the densely packed microtubule networks within neuronal processes, as physical expansion of the sample better allowed for isolated imaging of individual microtubules compared to STED imaging on a non-expanded sample (Katrukha et al., 2021).

Despite its accessibility, ExM has major downsides such as the time cost of imaging large, expanded samples and the fact that this is not a live cell compatible technology. ExM was also initially restricted to labeling proteins only. Several iterations of ExM have addressed this challenge and enabled labeling of other biomolecules, such as RNA (Chen et al., 2016), or lipids (Wen et al., 2020). Click-ExM (Sun et al., 2021), for example, is a new, versatile method that takes advantage of click chemistry to overcome biomolecule labeling limitations. The technique first involves labeling a biomolecule of interest with a clickable functional group such as an azide or alkyne. Next, the corresponding azide or alkyne biotin conjugate is attached via click chemistry. Finally, fluorescently labeled streptavidin binds this biotin and anchors it to the gel matrix before expansion. This technique is compatible with proteins, lipids, glycans, DNA, RNA, and small molecules.

Another major limitation of ExM is that initially it did not reach the ~20 nm resolution that SMLM or STED can offer. To get around this challenge, ExM protocols were developed that increase the expansion factor to improve resolution to ~25 nm. The first, termed iterative expansion microscopy (iExM) (Chang et al., 2017), involves two sequential ~4.5X expansions for an expansion factor of ~20X. The second, termed X10 microscopy (Truckenbrodt et al., 2018), expands the sample by a factor of ~10X in all directions in a single step by utilizing a new hydrogel recipe, though this expansion is more labor-intensive than other protocols and may not be suitable for thick tissue samples. Recently, a new protocol termed Ten-fold Robust Expansion Microscopy (TREx) was introduced, allowing for 10X sample expansion with a simple and reproducible protocol (Damstra et al., 2021). Another promising advancement was the introduction of Pan-ExM, which utilizes sequential expansions to achieve a 13–21X expansion factor combined with a global “pan-” proteomic fluorescent marker to achieve electron microscopy like structural contrast (M’Saad and Bewersdorf, 2020). This method offers added molecular context around a target protein of interest detected via normal immunofluorescence labeling. Together, these techniques reach close to the same resolutions as SMLM or STED, but have the advantage that they are more accessible, may offer improved visualization of densely packed proteins, and can achieve high resolution for low cost albeit only in fixed samples.

Combined and Correlative Approaches

ExM combined with super-resolution

Combining ExM with other super-resolution techniques is an emerging field, and the advancements mainly focus on improving the ~70 nm resolution of imaging after ~4–4.5X expansion of the sample. ExM has been combined with SIM, achieving a resolution of ~30 nm (Wang et al., 2018). Combinations of ExM with either SMLM or STED offer the exciting possibility of achieving even higher resolution, but also introduce unique challenges. During the expansion protocol, samples are treated with various harsh agents that either denature or digest proteins, resulting in loss of labeling epitopes. Because of this, most ExM protocols involve antibody labeling before expansion rather than after expansion. This approach presents further challenges, because even the best fluorophores lose a significant amount of fluorescence during the expansion process (Wassie et al., 2019). For SMLM or STED microscopy, labeling efficiency is crucial, and this loss during expansion can result in incomplete staining patterns or undercounting of targets due to signal quenching. This limitation was overcome by labeling targets of interest with multiple primary antibodies recognizing different epitopes to perform STED imaging, but this approach may not be feasible for most target proteins (Gao et al., 2018). Thus, for these combination techniques, post-expansion labeling is generally necessary. Post-labeling ExM has the distinct advantage that the linkage error between target protein and fluorescent label is reduced proportional to the expansion factor. With this advantage in mind, post-expansion labeling ExM was successfully combined with STORM to achieve ~4 nm resolution imaging of microtubules and centrioles, which retained their labeling epitopes after the expansion process (Zwettler et al., 2020).

While achieving this level of resolution is impressive, it was still limited to a subset of proteins that retain labeling epitopes after expansion. In 2021, Label-Retention ExM was introduced, which overcomes this problem using an optimized expansion protocol combined with SNAP and CLIP tags targeting proteins of interest (Shi et al., 2021). These small epitope tags are not degraded during the expansion process and are efficiently labeled with antibodies post-expansion. This technique, combined with STORM, allowed for ~4 nm resolution, and can be used to visualize any target that can be labeled with a SNAP or CLIP tag. This combination of techniques is highly promising, and we look forward to its use for future biological discoveries.

Light-Sheet and Lattice Light-Sheet microscopy combined with super-resolution

Light sheet microscopy addresses two major challenges of fluorescence imaging: imaging samples with greater depth and doing so with minimal photobleaching. Unlike confocal, the sample is illuminated from the side with a sheet of light such that sequential Z-planes can be imaged without illuminating the entire sample, dramatically reducing photobleaching. Lattice light sheet microscopy (Chen et al., 2014) utilizes the principles of light-sheet microscopy, but improves the technique by generating ultra-thin light sheets that illuminate an entire field of view with low peak excitation intensity, thus leading to high axial resolution, reduced photobleaching/phototoxicity and increased imaging speed.

LLS is highly compatible with super-resolution techniques, improving the penetration depth of these methods, which are typically limited to imaging thin specimens. LLS was first combined with both SIM and PALM (Chen et al., 2014). Recently, LLS combined with 3D-PALM was used to image the accessible genome in mouse embryonic stem cells, providing new insights into the organization of accessible regions of chromatin (Xie et al., 2020). LLS has also been combined with STORM to image the movement of plasma membrane receptors on whole live cells, highlighting the promise of combining high-speed 3D imaging with single-molecule precision (Waldchen et al., 2020).

Another exciting combination is LLS-ExM, which has enabled imaging an entire drosophila brain and mouse cortex with sub-cellular detail (Figure 4A). These techniques are theoretically quite compatible, as ExM physically expands samples for greater resolution, and LLS offers high-speed 3D imaging of large samples, but the technical challenges of combining them highlight some of the drawbacks of LLS microscopy. LLS requires highly specialized equipment, and uses a two-objective setup that requires specialized sample prep and sample mounting that generally are not compatible with standard cell culture imaging dishes and microscope slides. Several groups have attempted to create light-sheet microscopy setups that are more compatible with standard biological sample preparation. One particularly promising recent development is a single-objective oblique plane microscope with a high numerical aperture (Sapoznik et al., 2020). This setup offers comparable resolution to LLS microscopy for large, thick samples, while remaining compatible with standard biological sample preparation and mounting.

Figure 4: Combined and correlative approaches.

(A) Combined expansion and lattice light sheet microscopy showing individual neurons marked with YFP, as well as two individual synaptic proteins, Bassoon and Homer1, in a mouse cortical brain slice. Adapted from (Gao et al., 2019). (B) Correlative dSTORM and platinum-replica EM of dense core vesicle proteins Rab3a and Rab27a. Adapted from (Prasai et al., 2021)

Electron Microscopy combined with super-resolution:

Correlative light and electron microscopy (CLEM) has emerged as a powerful technique that combines the molecular-scale resolution and ultrastructural context of electron microscopy with the precise target specificity of light microscopy. However, without super-resolution imaging, CLEM is limited to targets that are either relatively large or sparsely distributed within cells. Combining super-resolution (SR) and EM (SR-CLEM), while potentially powerful, is an extremely difficult technical challenge. The EM and SR images must be properly aligned to one another, and since both techniques achieve very high resolution, a misalignment of only a few nanometers can be a major issue. SR-CLEM also requires using sample preparation methods that are compatible with both techniques, and research into how best to prepare samples and set up downstream imaging is ongoing. There have been several promising applications of SR-CLEM in recent years, using different strategies to overcome these technical challenges.

Correlative iPALM and metal replica transmission electron microscopy (TEM) utilizes gold nanorods coated in a thin layer of silica as a fiducial marker for alignment of fluorescence and TEM images. iPALM imaging is done first on targets of interest, and importantly gold nanorods can be visualized in this setup. Next, samples are prepared for TEM, and a metal replica is generated by coating the sample with a thin layer of platinum and carbon. Critically, when this replica is transferred from the sample and onto a TEM grid, the gold nanorods are embedded in the replica and can be visualized in TEM images. Sample warping can occur during this transfer process, adding another layer of technical difficulty. Nonetheless, the two sets of images can be aligned in X and Y using the position of the gold nanorods. One major limitation of this technique is that it relies on iPALM for Z-plane correlation of the fluorescence images to TEM images and is thus only applicable to targets at the plasma membrane of cells. However, it has been used to gain valuable insight into the organization of endocytic proteins at the plasma membrane with unprecedented resolution (Prasai et al., 2021; Sochacki et al., 2017) (Figure 4B).

SR has also been combined with cryogenic-electron tomography (Cryo-ET) in limited cases. The main challenge is achieving high quality SR images after cryo-preservation without disrupting the sample. The sample must remain under cryogenic conditions, while the objective lens for fluorescence imaging typically remains at room temperature. This necessitates the use of air-objectives at long working distances, which have much lower NA (~0.8) than oil immersion objectives (~1.3) used for SR imaging at room temperature. Additionally, while photobleaching is reduced, photoswitching is significantly slower at cryogenic temperatures than at room temperature, increasing the necessary imaging time. Further still, excitation laser intensity must be dramatically reduced, as the high laser intensity typically needed for SMLM can result in an increase in sample temperature, devitrifying ice. This too slows down photoswitching and results in longer acquisition times. These and other technical challenges of combining these techniques have been described in great detail recently (Dahlberg and Moerner, 2021). Despite these limitations, correlative PALM and Cryo-ET has been used to identify type VI secretion systems within Myococcus xanthus bacteria (Chang et al., 2014). Additionally, proof-of-concept of correlative 3D-PALM and cryo-ET has been demonstrated on thin cryosections of HEK293 cells (Liu et al., 2015). Using a custom setup, another group was able to achieve ~9 nm precision with a ~30 nm registration error between SR and cryo-ET images on three specific protein targets in Caulobacter crescentus (Dahlberg et al., 2020). Finally, an exciting application was the combination of focused ion beam scanning electron microscopy (FIB-SEM) with 3D cryogenic SMLM to image entire vitreously frozen cells (Hoffman et al., 2020). This imaging combination offers tremendous insight into the internal organization of human cells. However, it requires custom-built setups, and imaging for a single cell took 1–2 days per color for 3D SMLM, as well as 10–15 days for EM sample preparation and imaging.

Overall, as outlined above, SR-CLEM methods are extremely technically challenging, yet simultaneously extremely powerful. When implemented successfully, the payoff of combining precise target specificity with critical cellular context at molecular-scale resolution can be tremendous. Though this technique will likely not be accessible to the vast majority of researchers in the very near future, the wealth of information generated by SR-CLEM imaging, if shared freely, would allow researchers across different fields to examine their biomolecules of interest in their native context with new molecular-scale insight.

Conclusions and Outlook

Super-resolution microscopy has become an enabling tool for biological discovery in a short time but several challenges remain to be solved. Penetration depth of most commonly used and widely adopted super-resolution methods remains limited. While STED and RESOLFT, being point scanning methods, are more compatible with imaging thicker samples than the wide-field based SMLM methods, imaging very deep into tissues and intact organisms remains challenging. Combining super-resolution modalities with more specialized illumination geometries such as two-photon or light sheet in addition to tissue clearing or adaptive optics based corrections potentially hold promise for using these methods in more challenging biological settings (Bancelin et al., 2021; Chen et al., 2014; Hao et al., 2021b; Mlodzianoski et al., 2018; York et al., 2011). In particular, when combined with the quantitative and multiplexing capabilities of DNA-PAINT, such methods can advance our understanding of the molecular organization of cells, tissues and even intact organisms with high quantitative power. However, implementations of super-resolution microscopy with improved depth penetration are still a specialist’s tool due to being more technically challenging and remain to be widely adopted. Another promising avenue towards super-resolution imaging of thick biological samples is the combination of ExM with light sheet based approaches.

Super-resolution modalities, in their most commonly used form also remain limited to imaging fixed samples due to compromise between acquisition speed and spatial resolution, photobleaching and phototoxicity. RESOLFT and derivative approaches with their reduced photobleaching and phototoxicity combined with massively parallelized illumination can push the capability to image large fields of view at high spatiotemporal resolution, providing a promising path forward for both live and fixed cell super-resolution microscopy (Bodén et al., 2021; Chmyrov et al., 2013; Mahecic et al., 2019). Along similar lines, the imaging throughput of most super-resolution microscopy methods is low, making it challenging to acquire sufficient statistics even in fixed samples for robust biological interpretations. Massively parallelized illumination in STED/RESOLFT as well as new developments in SMLM that combine microfluidic sample handling, automated image acquisition and large field-of-view illumination are starting to address this challenge (Almada et al., 2019; Barentine et al., 2019; Beghin et al., 2017; Douglass et al., 2016; Mahecic et al., 2019; Stefko et al., 2018). However, most of these approaches are still in their infancy and remain to be widely adopted and more accessible.

The initial implementations of super-resolution microscopes, while drastically improving spatial resolution, did not reach true molecular scale resolution. While MINFLUX, MINSTED and related approaches are an important leap forward towards imaging at the molecular scale, more effort is needed for the development of bright, photostable, photoswitchable fluorophores as well as labeling approaches that minimize linkage errors and enable endogenous labeling. Development of new super-resolution compatible fluorophores and labeling methods has lagged behind the immense developments we have seen in optical microscopes, limiting the potential of these cutting-edge microscopes to reach true molecular scale resolution. Such improved labeling and sample preparation methods will also enable more robust combination of ExM with SMLM or STED based super-resolution microscopy, further pushing the resolution limit towards molecular scale imaging (Shi et al., 2021).

With advancements in microscopy and biological applications came the need for better quantitative analysis tools, in particular for SMLM methods that heavily rely on pre- and post-processing for image analysis. Several methods have been developed to not only generate/visualize super-resolution images but to also carry out segmentation, co-localization analysis and classification of morphologically diverse structures captured in super-resolution microscopy. While some efforts have been made to develop unified and modular analysis platforms for super-resolution data (Ries, 2020), most methods have remained disperse. There is a need for a concerted effort from method developers to consolidate the wide range of existing algorithms into a user-friendly, modular and robust platform that is easily accessible to new researchers entering the super-resolution field. In particular, incorporation of machine and deep learning approaches hold great promise for extracting novel biological insight from a diverse range of super-resolution images.

Overall, despite its recent history, super-resolution microscopy has been making a big impact in biology and it will be exciting to watch continued advancements in instrumentation, new fluorophores, labeling approaches and quantitative analysis tools, which will enable new biological applications. Commercialization of some of the more advanced modalities will be an important step towards accessibility of these tools to a wider range of researchers. In the future, we envision a multi-modal “smart microscope” that is operated in a fully automated mode, free of user input, and guided by the properties of a specific biological sample. Such a microscope could, for example, acquire lower resolution images with high throughput, analyze these images at high speed using neural networks, and use the analyzed information to perform a guided higher resolution imaging of a specific region of interest or upon a biological trigger. We are already starting to see examples of such “sample adaptive” and “smart” microscopes (Alvelid et al., 2021; Dreier et al., 2019; Gottfert et al., 2017; Heine et al., 2017) and this would be a particularly exciting field to watch.