Fluorescence microscopy has become an essential tool to study biological molecules, pathways and events in living cells, tissues and animals. Compared with other imaging techniques, such as electron microscopy, the main advantage of fluorescence microscopy is its compatibility with living cells, which allows dynamic and minimally invasive imaging experiments. The main weakness of fluorescence microscopy, however, has been its spatial resolution, which was limited to 
200 nm for many years.
At one end of the resolution spectrum, positron-emission tomography, magnetic resonance imaging and optical coherence tomography provide real-time read-out from animal or human subjects, but they cannot discern details smaller than 1 mm, 100 
m and 10 m, respectively (Fig. 1). At the opposite end, electron microscopy provides near molecular-level spatial resolution, but cells must be fixed, which is invasive and prevents dynamic imaging. Between these two resolution extremes, fluorescence microscopy provides a range of spatial and temporal resolutions. The most widely used fluorescence imaging methods, confocal microscopy and wide-field microscopy, can resolve certain cellular organelles (for example, the nucleus, the endoplasmic reticulum and the Golgi apparatus) and can track proteins and other biomolecules in live cells. The spatial resolution limit, however, prevents the resolution of single synaptic vesicles or pairs of interacting proteins. Many fields of biology would benefit from improved combinations of spatial and temporal resolution. For example, neuronal synaptic vesicles are 40 nm in size and signalling occurs on the milliseconds timescale. Bacteria are only 1-5 m in size, and subcellular features are difficult to resolve by conventional fluorescence microscopy.
Figure 1 | Comparison of the spatial and temporal resolutions of biological imaging techniques.
The size scale is logarithmic. Average sizes of biological features are given; specific sizes vary widely among different species and cell lines. The spatial and temporal resolutions are estimates of current practices, and some were taken from Ref. 120. The spatial resolution is given for the focal plane. The temporal resolution is not applicable (NA) for electron microscopy (EM) or near-field scanning optical microscopy (NSOM) because they image static samples. Ground-state depletion (GSD) and saturated structured-illumination microscopy (SSIM) have not been shown on biological samples, and thus their temporal resolutions are not determined (ND). ER, endoplasmic reticulum; MRI, magnetic resonance imaging; OCT, optical coherence tomography; PALM, photoactivated localization microscopy; PET, positron-emission microscopy; STED, stimulated emission depletion; STORM, stochastic optical reconstruction microscopy; TIRF, total internal reflection fluorescence; US, ultrasound; WF, wide-field microscopy.
We focus on the recent emergence of new far-field fluorescence imaging techniques, which theoretically have no limit to their spatial resolution. We describe the current limitations in terms of spatial and temporal resolution, and discuss how some of these might be overcome through improvements in fluorescent probe technology. In general, two classes of probes are used for super-resolution imaging: fluorescent proteins (FPs) and non-genetically encoded probes, such as organic small-molecule fluorophores and quantum dots. We describe the desired characteristics, strengths and weaknesses of each probe class. We also suggest future improvements to probe design and targeting, which might help to bring us closer to molecular-resolution imaging in live cells in real time.
Overcoming the diffraction limit
In 1873, Abbe observed that focused light always results in a blurred or diffracted spot, and the size of the spot places a fundamental limit on the minimal distance at which we can resolve two or more features1 (TIMELINE). This spot is commonly represented by the point spread function (PSF). The mathematical expression for Abbe's finding is that the resolution of a fluorescence microscope is limited to 
/2nsin
in the focal plane (xy) and 2/nsin2 along the optical axis (z), where is the wavelength of the light used and nsin is the numerical aperture of the lens. As most lenses have a numerical aperture of <1.5 ( is <70°), the theoretical resolution limit for cell imaging (where must be longer than 400 nm to minimize harm to cells) is 150 nm laterally and 500 nm axially2. In fact, this is the current resolution obtainable with a wide-field microscope.
Because the axial resolution in wide-field microscopy is much poorer than the lateral resolution, research in microscopy instrumentation from the 1970s to the 1990s focused on the improvement of axial resolution (TIMELINE). Initially, the use of confocal3 and multiphoton microscopy4 provided optical sectioning but did not significantly improve axial resolution. Later, I5M (Refs 5,6) and 4Pi (Refs 7,8,9) microscopies combined the apertures of two opposing lenses to allow the visualization of cellular structures with a large improvement in axial resolution — down to 100 nm — using wide-field and confocal set-ups, respectively. Although axial resolution improved by sixfold, lateral resolution remained unchanged. It was not until the 1990s that fundamentally new microscopies revolutionized the imaging field, for the first time breaking the lateral resolution diffraction limit. These techniques are collectively named super-resolution imaging techniques.
Near-field super-resolution imaging. In 1992, a super-resolution image of a biological sample was obtained for the first time10 (TIMELINE). Near-field scanning optical microscopy (NSOM) was originally proposed in 1928 by Synge11, and was first demonstrated in 1984 using visible radiation on test slides12, 13. NSOM overcomes the diffraction limit by removing the lenses and thus eliminates the need for focusing. Instead, the light passes through a small aperture that is positioned close to the sample (in the near-field zone), such that light cannot substantially diffract. The lateral resolution, determined by the diameter of the aperture, is typically 20-120 nm. Although NSOM has been used to study the nanoscale organization of several membrane proteins14, 15, imaging in the near-field is technically challenging. The aperture probe is difficult to make, and the need for feedback to maintain a constant distance from an irregular sample limits the speed of image acquisition. In addition, NSOM cannot be used for intracellular imaging. These factors have prevented widespread use of NSOM in cell biology.
Far-field super-resolution imaging. In contrast to NSOM, lenses are used in far-field microscopy, and they are placed at a distance from the sample. For super-resolution imaging, the key to overcoming the diffraction limit is to spatially and/or temporally modulate the transition between two molecular states of a fluorophore (for example, a dark and a bright state). Some techniques achieve super resolution by narrowing the PSF of an ensemble image of many fluorophores. These techniques include stimulated emission depletion (STED)16, ground-state depletion (GSD)17, and saturated structured-illumination microscopy (SSIM)18, 19 and its recent combination with I5M (I5S) (Ref. 20). Other super-resolution imaging techniques detect single molecules and rely on the principle that a single emitter can be localized with high accuracy if sufficient numbers of photons are collected21. These techniques include photoactivated localization microscopy (PALM)22, fluorescence photoactivation localization microscopy (FPALM)23 and stochastic optical reconstruction microscopy (STORM)24 (TIMELINE).
Cell biology imaged at super resolution
We focus our attention on STED, PALM, FPALM and STORM techniques because of recent reports that show the ability of these techniques to achieve super resolution in biological samples (for an in-depth review of other high- and super-resolution imaging techniques see Ref. 25). Several other techniques26, 27, 28, 29 have recently been developed but it is too early to assess their potential for biology.
Imaging an ensemble of molecules. STED was the first far-field super-resolution imaging technique to be applied to cell imaging30. To break the diffraction limit, STED uses spatially modulated and saturable transitions between two molecular states. Specifically, the sample is illuminated by two laser beams: an excitation laser pulse is immediately followed by a red-shifted pulse called the STED beam (Fig. 2Aa). Excited fluorophores exposed to the STED beam are almost instantly transferred back to their ground states by means of stimulated emission. This nonlinear (nearly exponential) de-excitation of the fluorescent state by the STED beam is the basis of breaking the diffraction limit in STED imaging. Although both laser pulses are diffraction-limited, the STED pulse is modified to feature a zero-intensity point at the focal centre and strong intensities at the spot periphery (creating, for example, a doughnut shape). If the two pulses are superimposed, only molecules that are close to the zero of the STED beam are allowed to fluoresce, thus confining the emission towards the zero. This effectively narrows the PSF (for example, to 65 nm in Fig. 2Aa), and ultimately increases resolution beyond the diffraction limit. To obtain a complete subdiffraction image, the central zero is scanned across the sample. Using this scheme, STED microscopy has achieved 20 nm resolution in the focal plane31, 32 and, recently, 45 nm resolution in all three dimensions33.
Figure 2 | Cellular features imaged by super-resolution techniques.
Aa | Point spread function (PSF) of stimulated emission depletion (STED) microscopy. The focal spot of excitation light (bright red) is overlapped with a doughnut-shaped red-shifted light (dark red), which quenches excited molecules in the excitation spot periphery. This confines emission to a central spot. Scanning this central spot (called the zero) across the sample results in a subdiffraction image. Ab | Synaptic vesicle movement imaged with STED. Synaptic vesicles were immunolabelled in live neurons, and the movement of each vesicle was individually recorded; the sum of 1,000 individual movie frames (
1–1,000) depicts the movement of various synaptic vesicles. Ba | Stochastic optical reconstruction microscopy (STORM). The fluorescence image is constructed from highly precise localization of single molecules. In each imaging cycle, all fluorescent molecules in the field of view are switched off by, for example, a strong red laser. Only a small percentage of them are then switched on (green light) such that their images do not overlap, and their emission is recorded (red light) and used to localize their positions (white crosses) with nanometre accuracy. Bb | Multicolour and three-dimensional (3D) STORM imaging. Conventional (left panel) and STORM (right panel) images of immunostained microtubules (green) and clathrin-coated pits (red) in the same region of a BSC-1 cell. A 3D STORM image of a clathrin-coated pit is inset. An xy cross-section and an xz cross-section of the pit are shown in perspective. Ca | Photoactivated localization microscopy (PALM). PALM follows the same principle as STORM. To perform two-colour PALM imaging, the orange emitters (Eos fluorescent protein) are sequentially activated (405 nm light), imaged (561 nm light), localized and bleached until a subdiffraction image can be constructed (steps 1–3). After bleaching the remaining Eos molecules (step 4), the many active green emitters (Dronpa fluorescent protein) are first deactivated with a strong 488-nm light (step 5). Then, the green emitters are activated, imaged and bleached (steps 6–8). Cb | Two-colour PALM images show the nanostructural organization of cytoskeletal actin (green) and the adhesion protein paxillin (red) in an HFF-1 cell. Actin bundles densely cluster around some (arrowheads) but not all (full arrows) paxillin adhesions. Images in part Aa modified, with permission, from Ref. 32 © (2007) The Biophysical Society. Images in part Ab modified, with permission, from Ref. 40 © (2008) American Association for the Advancement of Science. Part Ba modified, with permission, from Nature Methods Ref. 24 © (2006) Macmillan Publishers Ltd. All rights reserved. Images in part Bb modified, with permission, from Ref. 46 © (2007) American Association for the Advancement of Science. Inset image in part Bb modified, with permission, from Ref. 49 © (2008) American Association for the Advancement of Science. Parts Ca, Cb modified, with permission, from Ref. 48 © (2007) National Academy of Sciences.
Since its invention in 1994, STED has been applied to several cell biological problems. STED resolved synaptotagmin-I in individual synaptic vesicles (40 nm in size), and showed that this protein forms isolated clusters upon vesicle fusion34. STED also revealed the ring-like structure of the protein bruchpilot at synaptic active zones in the Drosophila melanogaster neuromuscular junction35, and the size and density of syntaxin-I clusters in PC12 cells36. Additionally, STED has enabled the visualization of the nuclear protein splicing component-35 (SC35)31, the nicotinic acetylcholine receptor37, the transient receptor potential channel M5 (TRPM5)38 and the flotillin-2-induced clusters of the amyloid precursor protein39. Recently, STED was extended to two-colour imaging, enabling colocalization studies of two mitochondrial proteins32. This study required nanoscale-resolution imaging because the entire mitochondrion is only 200-500 nm.
Impressively, Westphal and colleagues reported video-rate imaging of synaptic vesicles with 62 nm lateral resolution (Fig. 2Ab) in live hippocampal neurons40. Using STED microscopy, the synaptic vesicles, labelled with ATTO647N-conjugated anti-synaptotagmin antibodies, were observed to be highly restricted inside synaptic boutons. By contrast, vesicles outside boutons exhibit faster, linear movement, which might represent transit through axons (Fig. 2Ab). Time-lapse imaging of live mammalian cells with <50 nm lateral resolution was also recently achieved using the yellow FP citrine targeted to the endoplasmic reticulum41. In this report, the use of a FP instead of an antibody label enabled live-cell imaging of intracellular features.
Although many applications of STED have been reported, the instrumentation required is still complicated and this is the main limitation to its widespread use. This limitation has been acknowledged and STED microscopy using continuous wave beams was recently reported42. This has the potential to be implemented in any conventional confocal set-up. In addition, the use of a commercially available super-continuum laser system eliminates the need for two laser sources43.
Imaging single molecules. PALM, FPALM and STORM are fundamentally different from STED in that they image single molecules. The basic principle behind these techniques is that the position of a single molecule can be localized to 1 nm accuracy or better if enough photons are collected and there are no other similarly emitting molecules within 200 nm. This notion, which was described by Heisenberg in the 1930s and mathematically formulated in the 1980s44, has enabled single-particle tracking studies, such as the imaging of the movement of kinesin-coated beads with a precision of 1-2 nm by Gelles et al.45, and is the basis of many other single-molecule imaging studies.
Localization accuracy does not, however, directly translate into super-resolution images, especially for densely labelled samples; the overlapping images of these fluorophores would prevent their accurate localization. To overcome this, super-resolution imaging techniques determine the nanoscale localization of individual fluorescent molecules by sequentially switching them on and off using light of different wavelengths (Fig. 2Ba). In contrast to STED, in which the switching is predefined in space through the superimposition of the two laser beams, the switching is done stochastically in single-molecule-based super-resolution methods. In each imaging cycle, most molecules remain dark, but a small percentage of molecules are stochastically switched on, imaged and then localized. Repeating this process for many cycles allows the reconstruction of a super-resolution image.
Such super-resolution techniques have been used to image protein–DNA complexes in vitro24 and to image molecular structures, such as lysosomes, mitochondria and adhesion complexes22, and microtubules and clathrin coated pits46 in fixed cells. Single-molecule-based super-resolution microscopy has also been extended to multicolour imaging46, 47, 48. Two-colour STORM revealed the organization of microtubule networks and clathrin-coated pits in fixed mammalian cells with 20 nm resolution46 (Fig. 2Bb). Using PALM (Fig. 2Ca), two-colour imaging of actin and adhesion complexes in fixed cells was reported48 (see Fig. 2Cb, which shows fibrillar-like adhesions of paxillin running parallel to actin fibres). At nanoscale resolution, little overlap is observed between actin and paxillin, although actin bundles densely cluster around some but not all paxillin adhesions (Fig. 2Cb). This structural relationship could not be viewed using conventional microscopy.
Three-dimensional (3D) super-resolution imaging has also been achieved with STORM and FPALM. Using optical astigmatism, Huang and co-workers performed STORM imaging with 20-30 nm resolution in the xy plane and 50 nm resolution in the axial dimension49. Using multifocal plane imaging, Juette and co-workers achieved 3D FPALM imaging with 30 nm resolution in the xy plane and 75 nm resolution in the axial dimension50.
Recently, PALM and FPALM have been extended to live-cell imaging51, 52, 53. For example, Hess and colleagues imaged the distribution of the membrane protein haemagglutinin from influenza virus in live fibroblasts with 40 nm accuracy, and determined an effective diffusion coefficient of 0.1 m per second51. Manley et al. imaged individual tsO45 vesicular stomatitis virus G particles and HIV-1 Gag membrane proteins in living cells, and obtained high-density molecular tracks52. Shroff and colleagues imaged adhesion-complex dynamics in live CHO cells with 60 nm resolution at an imaging speed of 25–60 seconds per frame53. They observed the migration of adhesion complexes away from the cell edge, which had previously been seen by conventional microscopy. However, super resolution allowed them to observe, for the first time, that the rear end of migrating adhesion complexes moves faster than the front. Also, new adhesion complexes form in the cell interior rather than at the cell edge.
Limitations of super-resolution imaging
Super-resolution microscopy has improved both lateral and axial resolutions, recently achieving 50-70 nm resolution in all three dimensions33, 49, 50. However, we are far from molecular resolution (1–5 nm), in which individual molecules in a macromolecular assembly can be resolved. What factors determine the spatial resolution limit for each of these methods?
Ensemble super-resolution imaging. In STED, the resolution depends on the extent of the saturation of emission depletion, because this defines the degree to which the PSF can be narrowed (Fig. 2Aa). This relationship is given by equation 1:
In this equation, is the wavelength, nsin is the numerical aperture of the microscope, Imax is the applied intensity of the STED pulse and Isat is the STED intensity that gives 50% depletion of the emission31. It therefore follows that to maximize the saturation of emission depletion (Imax/Isat) and improve STED resolution, one needs to either increase the intensity of the STED pulse (increase Imax) or decrease the intensity needed to send a particular fluorophore to the dark state (decrease Isat). For example, a typical Imax intensity used in early STED studies of 250 MW per cm2 produced a lateral resolution of 50-70 nm for the fluorescent small molecule ATTO532 (Ref. 31). Increasing Imax above this value was not possible because the probe photobleached too quickly. Later, the use of STED pulses of longer duration (to reduce ground-state multiphoton absorption)54 and of lower frequency (to allow the triplet state to relax and avoid triplet-state excitation)31 resulted in a marked decrease in fluorophore photobleaching. These improvements enabled the increase of Imax to 2,200 MW per cm2 and resulted in 15–20 nm resolution for the same dye31.
Because there will always be an upper limit to Imax that is imposed by probe photobleaching and sample damage, an alternative approach is to decrease Isat (a separate characteristic of each fluorophore). Isat defines the intensity at which the rate of stimulated depletion is faster than other competing interstate transitions, such as fluorescence emission or excitation to higher (singlet or triplet) excited states, and Isat is inversely proportional to both the fluorescence lifetime of the fluorophore and its cross-section for emission depletion. Thus, good STED dyes are characterized by high quantum yields, emission spectra that match the STED wavelength, enhanced photostability, long fluorescence lifetimes (>0.8 ns)55 and a low cross-section for multiphoton absorption and for absorption by the excited states.
Another approach to decrease Isat is to change the nature of the 'bright' and 'dark' states of the probe. We described above a stimulated depletion that brings the molecule from the excited state S1 (bright) to the ground state S0 (dark), but STED can also work if, for example, the two states are two different molecular states of a photoswitchable probe. This more general approach to super-resolution imaging is termed reversible saturable optically linear fluorescence transitions (RESOLFT)56, which applies to all ensemble techniques based on a stimulated transition between any two molecular states (for example, STED, SSIM and GSD). Because the spontaneous interstate transition is almost non-existent when using photoswitchers, Isat is much smaller and therefore the resolution can be improved, even with low laser intensities. This was first shown by Hofmann et al.56, who obtained 50-100 nm resolution in the focal plane using the photoswitchable protein FP595 (isolated from Anemonia sulcata; Box 1) with a STED power of only 100-600 W per cm2, which is six orders of magnitude lower than that used for ATTO532 (Ref. 31), and similar to the intensities used, for example, in FPALM imaging51.
It is also advantageous to maximize the speed of super-resolution imaging, to allow the study of dynamic processes in living cells. Initial acquisition times for PALM, FPALM, STED and STORM super-resolution techniques were of the order of minutes to hours, which restricted them to the imaging of fixed samples. Recently, STED was applied to imaging of synaptic vesicle movements in living cells at video rate, or 28 frames per second40. The imaging speed in STED is restricted by the minimum number of photons that can be collected per pixel and per unit time. The high imaging speed in STED was obtained at the cost of increasing laser intensity (to 400 MW per cm2) and reducing the number of photons collected per imaging cycle, which caused a reduction in spatial resolution (to 62 nm) as well as field of view (to 2.5 m 
1.8 m). The goal is to achieve this same video-rate imaging speed while maximizing spatial resolution and using laser intensities that are more appropriate for living cells.
Single-molecule-based super-resolution imaging. In PALM, FPALM and STORM, spatial resolution is determined by the number of photons that can be collected from each fluorophore and by the background fluorescence, as given by equation 2:
In this equation, 
x is the localization precision; k1 and k2 are determined by the excitation wavelength, the numerical aperture of the objective and pixel size; N is the number of collected photons (a fraction of the total emitted photons); and b is the background noise per pixel21. Thus, to maximize resolution, the aim is to minimize background noise and maximize photon output of the fluorophore. For example, in the absence of background, if 10,000 photons can be collected from a single fluorophore molecule before it bleaches or is turned off, its localization can be determined to 2 nm precision, and 400 photons can provide 10–20 nm localization accuracy57. Background can arise from sample autofluorescence, as well as from residual fluorescence of surrounding probe molecules in the dark state. For single-molecule-based super-resolution imaging, it is then desirable for fluorophores to have a high contrast ratio, which is defined as the emission intensity ratio between the bright versus dark states. This is because at low contrast ratios, the collective fluorescence from dark molecules can obscure the signal from the small number of bright molecules during each imaging cycle. One way to reduce background is by using a total internal reflection fluorescence (TIRF) microscope, but this restricts imaging to the cell membrane. To maximize spatial resolution, it is also important to maximize the signal from fluorophores in the bright state. Thus, brighter fluorophores with high extinction coefficients and high quantum yields are desirable.
This need for low background and high photon output highlights some of the main differences between single-molecule and ensemble read-out schemes. A main advantage of the single-molecule-based super-resolution strategy is that the marker molecules are not forced to undergo several photoswitching cycles — this is the case for ensemble-based super-resolution imaging, in which photobleaching is a major problem. However, the fluorescent 'on' state must produce enough photons to allow its precise localization (which is predetermined in STED imaging). In addition, the single-molecule approach requires strict control over the maximum density of photoactivated molecules, and depends on their reliable localization against a diffuse background.
Spatial resolution should be improved without sacrificing temporal resolution. Live-cell PALM was recently used to study adhesion-complex dynamics53, but this was made possible by the slow intrinsic motion of adhesion complexes. At an imaging rate of 25-60 seconds per frame, many other biological movements would appear to be blurred. To improve the temporal resolution of PALM, FPALM and STORM, it is necessary to maximize the number of photons that can be collected per unit area per unit time. Manley et al.52 reported that high-density single-particle tracking with PALM was best achieved using EosFP, which has the largest contrast ratio and highest photon output of all the known photoshiftable FPs22, 48.
Another consideration is that when using irreversible fluorophores, temporal resolution is also limited by the photobleaching rate. In these cases, less photostable probes are desired, although a balance must be met between faster photobleaching and adequate photons to achieve high localization accuracy. In other words, irreversible probes should be very bright but not necessarily photostable. The alternative is to use reversible fluorophores such as cyanine (Cy) dyes58 or the PA-FP Dronpa59, which can be turned off and thus do not have to be photobleached.
FPs for super-resolution imaging
Although simple FPs, such as green FP (GFP) and yellow FP (YFP), have been used in STED imaging41, 55, most super-resolution imaging techniques exploit the intrinsic ability of certain FPs to change their spectral properties on irradiation with light of a specific wavelength. There are two main classes of FPs used in super-resolution imaging: those that convert from a dark state to a bright fluorescent state (PA-FPs), and those that change fluorescence wavelength on irradiation (photoshiftable FPs (PS-FPs)) (reviewed in Refs 60,61) (Box 1). All known PS-FPs irreversibly shift their wavelength but PA-FPs can photoactivate either reversibly or irreversibly.
Desired characteristics. FPs for super-resolution imaging should be as bright as possible (that is, have large extinction coefficients (
abs) and large fluorescence quantum yields (
fl)), to maximize the number of detectable photons per molecule (N) and they should have high contrast ratios. Additionally, the spontaneous interconversion rates into and out of the activated (fluorescent) state must be low compared with the light-controlled activation rate. The rates of photobleaching (for irreversible PS-FPs) or deactivation (for reversible switchers) should be balanced with the activation rate to ensure that, only a small number of molecules are in the fluorescent state at any given time, so that on average they are separated by more than the diffraction limit, and also to ensure that each activated molecule remains in the fluorescent state for long enough to give sufficient photons for accurate localization. Finally, for cellular studies, the FP should be monomeric to minimize perturbation of the target protein. As seen below, every currently available modulatable FP has at least one drawback.
Irreversible PA-FPs. Two irreversible PA-FPs have been engineered: PA-GFP and monomeric PA-RFP1-1. PA-GFP was the first to be engineered, developed by mutagenesis of the original GFP62. Although PA-GFP has been used in FPALM imaging to determine the diffusion coefficient of haemagglutinin in live fibroblasts51, its green fluorescence and low contrast ratio results in high background, which limits spatial resolution to 40 nm and necessitates a minimum acquisition time of 150 ms per frame (about sevenfold slower than the acquisition time required for rsFastLime (also known as Dronpa V157G), a variant of the reversible PA-FP Dronpa63). The low contrast ratio and extremely low quantum yield of the only irreversible red PA-FP, monomeric PA-RFP1-1 (Ref. 64) (Box 1), which is derived from DsRed, have thus far prevented its use in super-resolution imaging.
Irreversible PS-FPs. Many of the naturally occurring and engineered PS-FPs exhibit a shift from green to red emission (Box 1). Among these, the natural PS-FP kaede65 and the engineered KikGR66 are obligate tetramers, and are thus not suitable for imaging of cellular proteins. EosFP is the most commonly used red PS-FP for super-resolution imaging, as it has the highest contrast and brightness and has been engineered into a monomeric form that is suitable for protein fusion (Eos; Box 1)67. EosFP was used in one of the first demonstrations of PALM imaging22, and it was later used in combination with PS-CFP2 and Dronpa in two-colour PALM imaging experiments48. It was because of the optimal photophysical properties of EosFP that Manley and co-workers could perform single-particle tracking of membrane proteins in live COS7 cells at an imaging speed of 20 frames per second52. EosFP in its original dimeric form was also used in the latest demonstration of PALM in live cells53. The main disadvantage of monomeric Eos, however, is that chromophore formation occurs only at temperatures below 30°C, which limits its use in mammalian cells67. In this regard, Dendra-2, with similar contrast and brightness, could potentially outperform monomeric Eos, because it matures properly at 37°C and can be activated by blue light (which causes less damage to tissue than the ultraviolet (UV) light that is required for monomeric Eos)68. However, even the brightest PS-FP is still much dimmer than some small-molecule organic fluorophores. For example, Eos provides 490 collected photons per molecule48, whereas the switchable fluorophore pair Cy3–Cy5 provides 6,000 collected photons per molecule per switching cycle and lasts 200 switching cycles46, 58.
The only green-emitting PS-FP, PS-CFP2 (Ref. 69), is preferred in multicolour studies because it has the highest contrast ratio and yields the largest number of photons of all of the green modulatable FPs48 (Box 1).
Reversible PA-FPs. Reversible PA-FPs (also known as photoswitchers) are advantageous in super-resolution imaging because the same fluorophore can be imaged multiple times. Reversible photoswitching is mandatory in RESOLFT imaging, in which each molecule is switched on and off many times in order to reconstruct a subdiffraction image.
The first reversible PA-FP reported was FP595 (Ref. 70). Although FP595 has low contrast and is tetrameric, it was successfully used by Hofmann et al. to achieve 50–100 nm focal plane resolution using RESOLFT imaging56. The best-known reversible PA-FP is the naturally occurring Dronpa59 and its many engineered variants. Unfortunately, although Dronpa exhibits a large extinction coefficient and quantum yield, its fluorescence is excited with 488 nm light, which also inactivates the protein, therefore resulting in a low number of collected photons per imaging cycle. Higher spatial resolution can be obtained with PS-CFP2 than with Dronpa48.
To overcome this limitation, Andresen and co-workers recently engineered a Dronpa variant with positive-switching characteristics71. In contrast to Dronpa, this new variant, called Padron, is activated and imaged by blue light, whereas UV light switches the protein off. It remains to be seen how the protein behaves for super-resolution imaging. As an additional example of how probe development combined with instrumentation advances can lead to improvements in imaging, Egner and co-workers have recently shown that the use of a faster photoswitching variant of the FP Dronpa, rsFastLime72, in combination with asynchronous recording, accelerates imaging acquisition and eliminates the need for TIRF in PALM. The intense light used in this setting drives most of the fluorophores into a dark state. Individual fluorophores then stochastically and spontaneously revert to the bright state, briefly emit a burst of photons, then revert to the dark state. In this scheme, the acquisition time matches the mean duration of an emission burst, producing rapid acquisition and lower background fluorescence. This procedure has been termed PALMIRA (PALM with independently running acquisition)47, 63, 73.
Finally, the recent engineering of the first monomeric red reversible PA-FP, rsCherry, has opened new possibilities for multicolour time-lapse imaging. Live-cell PALMIRA imaging of the endoplasmic reticulum labelled with rsCherry provided images with 75 nm lateral resolution74. As described for Padron, the authors have also engineered a positive-switching version of monomeric cherry, termed rsCherryRev, which might further enhance its use, although its applicability to cell imaging has not been shown74.
In summary, FPs can be targeted with absolute specificity but they are generally bigger, dimmer and less photostable than small-molecule fluorophores. This low brightness usually makes it necessary to use a TIRF microscope to minimize background fluorescence. Even with the brightest PS-FP, EosFP, the maximum frame rate that can be achieved (25 seconds per frame) is still too slow to image most biological processes if a resolution of 60 nm is desired. Brighter FPs are needed to increase temporal resolution without losing spatial resolution. Unfortunately, although the mechanisms of photoswitching for some FPs have been recently described61, the strict requirement for chromophore formation inside the 
-barrel of FPs makes the engineering of brighter FPs a difficult task and highlights the need for the discovery of new FPs from other species. Additionally, new monomeric proteins of different colours are needed to allow routine multicolour imaging at super resolution.
Non-genetically encoded probes
Three main classes of non-genetically encoded probes have been used in super-resolution imaging: inorganic quantum dots (Box 2), reversible photoactivatable fluorophores (also known as photoswitchers) and irreversible photoactivatable fluorophores (also known as photocaged fluorophores) (Box 3). STED imaging initially used regular small-molecule dyes, in which the probe was imaged in its excited state and the surrounding molecules were quenched by the STED pulse that sent them to ground state. In this type of imaging scheme, the intensity of the STED pulse needs to be extremely high to compete with the spontaneous fluorescence decay of the probe — molecules with both high quantum yield and long fluorescence lifetime (slow fluorescence decay), such as the ATTO or DY dyes, are ideal30, 31, 32, 34, 35, 36, 37, 38, 40, 42 (Box 3). Later, STED imaging evolved into the more general RESOLFT imaging, which uses photoswitchers, such as FP595 (Ref. 56), and furyl fulgides75.
Reversible PA probes. The small-molecule analogues to reversible PA-FPs (such as Dronpa and rsCherry74) are photochromic probes, including rhodamines and diarylethenes, as well as photoswitchable cyanines. The switching mechanism of a photochromic rhodamine B (PC-RHB) is depicted in Box 3. Irradiation of the closed isomer with UV light or with red light (for pulsed two-photon absorption) results in transient formation of a coloured and brightly fluorescent open isomer73. The reaction is thermally reverted by heat within milliseconds to minutes, depending on the solvent. Irradiation of the fluorescent isomer with green light excites fluorescence emission but does not regenerate the non-fluorescent state. This property makes the photochromic rhodamine superior to Dronpa (similar to Padron and rsCherryRev), because the fluorescence signal can be read without the undesired erasing side effect, which results in higher photon output per switching event. Photochromic rhodamines also provide an example of how probe development can lead to improvements in the imaging process. Folling and co-workers76 reported the design of a new probe based on rhodamine 590 (PC-RH590), whose extra rigidity improves cross-section compared to the original PC-RHB73 for two-photon absorption. Efficient two-photon activation could also be achieved with this probe, and fluorescent images with 15 nm resolution in the focal plane were obtained76.
Other photochromic molecules, such as photochromic diarylethenes77, could potentially be used in super-resolution imaging, but they have limited water solubility, which restricts their biological utility.
Photoswitchable cyanine dyes have been used in both PALMIRA47 and STORM24, 46, 49 imaging. Although it has also been used alone47, Cy5 is best used in combination with a secondary chromophore that facilitates the switching46, 58. For example, when Cy5 is paired with Cy3, the same red laser that excites Cy5 is also used to switch the dye to a stable dark state. Subsequent exposure to green laser light converts Cy5 back to the fluorescent state, and this recovery rate depends on the close proximity of the secondary dye Cy3 (called the activator)58. Cy5 switching can also be facilitated by other activator fluorophores, such as Alexa Fluor 405 and Cy2 (Ref. 46). Furthermore, Cy3 was found to facilitate switching of other cyanines, such as Cy5.5 and Cy7 (Ref. 46). This finding greatly increases the palette of colours that are available for STORM imaging and has allowed simultaneous visualization of microtubules and clathrin-coated pits in fixed mammalian cells with 20-30 nm lateral resolution46 (Fig. 2Bb). This availability of several colours contrasts with the lack thereof for modulatable FPs. Unfortunately, the development of improved switching pairs is currently hindered by the fact that the switching mechanism of these dyes is unknown. Finally, the recently developed Cy3–Cy5 double conjugate facilitates labelling78.
When compared with their FP counterparts (FP595, Dronpa and Padron; Box 1), photoswitchable dyes have larger contrast ratios, and they have higher extinction coefficients, which results in brighter probes with higher numbers of collected photons per molecule. Of the photoswitchable probes, rhodamines stand out because of their potential for intracellular labelling in live cells, as they are membrane permeable (sulphonated cyanine dyes are not). Improvements are required, however, to reduce the known affinity of rhodamines for intracellular organelles that is caused by their hydrophobicity and positive charge.
Irreversible PA probes. Caged compounds, such as caged Q-rhodamine79, 80, can also be exploited in super-resolution imaging (Box 3). During uncaging, irradiation with UV light causes the release of a protective group and results in a large increase in the fluorescence intensity of the dye. Photocaged probes can be used in PALM, FPALM or STORM in the same way as irreversible PA-FPs: they can be uncaged, localized with high precision and then bleached. Caging can be a way of generating new PA probes that are based on fluorophores with otherwise good photophysical properties but without intrinsic photoswitching ability. The potential of caged compounds was first shown by Betzig and co-workers, who used PALM to image caged rhodamine–dextran that had been dried on glass cover slips22. Caged fluorescein has also been used to image beads, achieving sub-100 nm resolution in all directions50. A new irreversible photoactivatable compound based on a dicyanodihydrofuran fluorophore has been developed by Lord et al.81, who reported photoactivation and live-cell imaging of the dye in mammalian cells, although not with super resolution. Caged compounds have not yet been used for super-resolution imaging of biological samples.
Targeting methods for non-genetically encoded probes. Although non-genetically encoded probes generally show increased brightness and photostability compared with FPs, they also have disadvantages. The lack of genetic encoding means that these probes require a means to target them to the biomolecule of interest inside cells. These probes have been traditionally targeted using antibody conjugation, although this has many disadvantages. Antibodies are not membrane permeable, and hence are not useful for intracellular labelling of living cells. Antibody staining also usually results in low labelling efficiency and the large size of antibodies adds uncertainty (10-20 nm) to the spatial relationship between the label and its target. Why does labelling efficiency matter? The label–target relationship was recently described by Shroff and co-workers by analogy with the Nyquist criterion; the distance between two labelled molecules must be smaller than half of the desired spatial resolution53. If we want a spatial resolution of 5 nm, we need to have labelled molecules at least every 2 nm. Given the bulk and weak affinity of many antibodies, it is difficult to achieve such labelling density. Another disadvantage of many non-genetically encoded probes is their inability to permeate through the cell membrane, which restricts their use to either fixed cells or cell-surface proteins.
In order to fully realize the intrinsic imaging resolution of super-resolution techniques, we need to develop strategies to target small molecules without significantly increasing the size of the tag and with high labelling efficiency. Box 4 describes some of the current approaches for site-specific protein labelling in live cells. In a subset of these methods, the protein of interest is fused to a peptide or protein sequence that recruits the small molecule. Examples of methods that use a peptide tag include the tetraCys82, 83, 118, hexaHis84, 85, 86, the polyAsp87, 88, 89 and the bungarotoxin-binding peptide90, 91 methodologies. Some other methodologies use proteins to recruit the small-molecule probe, such as the FKBP12 protein92 and the enzymes dihydrofolate reductase (DHFR)93, 94, O6-alkylguanine-DNA alkyltransferase (AGT)95, 96, 97, cutinase98 and dehalogenase99. The use of protein tags, instead of peptide tags, can improve the specificity of the binding owing to the larger interaction surface area that they can establish with the recruited probe, but the cost is the increased size of the tag, which can perturb protein function.
To bridge the requirements of high labelling specificity and minimal perturbation of the target protein, a second set of methodologies uses a peptide recognition sequence but then uses an enzyme to catalyse the covalent attachment of the probe to the peptide (Box 4). Using an enzyme to catalyse probe ligation improves the specificity of the labelling and also gives faster and covalent attachment of the probe. Examples of this second approach include the methods based on the enzymes sortase100, 101, transglutaminase102, biotin ligase103–105, phosphopantetheine transferases (Sfp and AcpS)106 and lipoic acid ligase107, 108.
Conclusions and future perspectives
These are exciting times in cell biology because live-cell imaging with molecular resolution (1-5 nm) is now closer to becoming a reality. The recent development of super-resolution imaging techniques has enabled the visualization of cellular features with previously unimagined detail. STED imaging has allowed real-time tracking of single synaptic vesicles in cultured hippocampal neurons40 and has revealed the anatomy and dynamics of syntaxin-I clusters in PC12 cells36. STED, STORM, PALM and FPALM allowed cellular structures to be imaged in 3D and in multiple colours32,33,46-50,109.. At such an improved resolution, colocalization of protein pairs has been reported that contradicts previous reports based on low-resolution imaging48, and diffusion properties of membrane proteins have been mapped to high resolution51, 52.
We have described the technological advances introduced by the STED, PALM, FPALM and STORM techniques, which have made far-field super-resolution imaging possible. We have also pointed out the future developments that we think are required to further improve the spatial and temporal resolution of these methods. Although improved computational methods and imaging equipment are also needed, fluorescent probes do limit performance of super-resolution imaging. In the future, we expect that smaller, brighter, more photostable and membrane-permeable fluorescent probes will allow video-rate imaging with molecular resolution.
