Labelling strategies for correlative light electron microscopy

Imaging is one of the key technologies underpinning discoveries in biomedical research. Each imaging technique however usually only provides a specific type of information. For instance, live‐cell imaging using fluorescent tags can show us the dynamics of a system. On the other hand, electron microscopy (EM) gives us better resolution combined with the structural reference space. By applying a combination of light and electron microscopy modalities to a single sample one can exploit the advantages of both techniques in correlative light electron microscopy (CLEM). Although CLEM approaches can generate additional insights into the sample that cannot be gained by either technique in isolation, the visualization of the object of interest via markers or probes is still one of the bottlenecks in a Correlative Microscopy workflow. Whereas fluorescence is not directly visible in a standard electron microscope, gold particles, as the most common choice of probe for EM can also only be visualized using specialized light microscopes. In this review we will discuss some of the latest developments of probes for CLEM and some strategies how to choose a probe, discussing pros and cons of specific probes, and ensuring that they function as a dual modality marker.


| INTRODUCTION
Fundamentally, correlative techniques revolve around imaging the same sample/region of interest with several complementary techniques. Different imaging modalities will be used to gather specific information based on the sample and required information. Data sets generated using a correlative methodology will often provide information not available to the individual constituent techniques. One of the most widely used correlative imaging workflows is correlative light electron microscopy (CLEM). CLEM uses light and electron microscopy techniques in tandem to reap the benefits of both while compensating for shortcomings in each (Ando et al., 2018, Verkade, 2021. The sum of such an experiment should provide more data/insight than each technique alone (hence we have termed this 1 + 1 = 3, Collinson & Verkade, 2015). The developmental advances in fluorescent marker proteins have enabled fluorescence microscopy to become molecularly sensitive and capable of observing target molecules in a living state. However, even with the super-resolution fluorescence microscopy technologies currently available, one of the limitations of fluorescence microscopy is its resolution mainly due to the diffraction limit (Kobayashi et al., 2016). Also, one will only visualize the marked fluorescent proteins and be unable to capture the subcellular context, unless additional fluorescent markers are used to mark cellular organelles/structures. By capturing a specific moment in time by fixation, these disadvantages of fluorescence microscopy: lower resolution, lack of ultrastructural context and relatively poor optical sectioning are redressed by imaging the same cell/location again with electron microscopy (EM) which excels in these areas (Ando et al., 2018;De Boer et al., 2015;Verkade, 2008;Verkade, 2021). Similarly, the inclusion of Fluorescence Microscopy  . There are many ways to perform a CLEM experiment and a wide variety of microscopy modalities that can be combined. An example of a CLEM workflow is shown in Figures 1 and 2. The main driver in instrument choice and the experimental approach should depend primarily on the posed scientific question (Verkade, 2021).
A CLEM experiment critically relies on being able to find the same cell/location in both techniques so that images and data generated in each imaging modality can be correlated. However, finding a specific event back in the electron microscope has always been one of the most difficult and time-consuming aspects of a CLEM experiment , Schmidt et al., 2021. To do this effectively, the objects of interest must be labeled with specialized probes or markers (Collinson & Verkade, 2015). Strictly speaking, a distinction between the 2 (marker and probe) could be drawn. A marker can be defined as already present in the sample (e.g., expression of a protein with a Green Fluorescent Protein [GFP] tag), whereas a probe is added to the exterior of the cell/tissue (e.g., fluorescent antibody labelling).
Generally, these terms are used interchangeably to denote a moiety F I G U R E 1 Example of a CLEM experiment using DAB precipitation of DAB. Hela cells were grown in Mattek finder dishes (a, see also 44). The cells were transiently transfected so some cells express the Epidermal Growth Factor receptor (EGF-r) tagged with GFP (b). Cells were incubated for 30 min with a 1:1 mixture of EGF-ALEXA 594 and EGF-HRP (c) and bright field images were acquired (d, e showing merge of b, c, and d). The coordinates of the cell of interest (arrowhead) were recorded using the coordinate system (arrow pointing to 8 in Figure f). Following processing for EM the finder pattern is imprinted into the resin (g) and after sectioning the cell of interest can be retraced in the TEM (H). Scalebar is 50 μm for (b-e), 100 μm for (f), 150 μm for (g), and 20 μm in (h). that allows specific identification of an object of interest (see e.g., Brown et al., 2012;Reed et al., 2013) We have decided not to make a distinction but will generally use the term probe throughout to describe both.

| ENDOGENOUSLY EXPRESSED PROBES
Since the discovery of GFP, a wide selection of cloneable fluorescent reporter molecules have become powerful tools in the microscopist's toolbox (e.g., Grimm & Lavis, 2022). These single modality genetic probes (mostly derived from GFP) provide one-to-one reliable targeting for LM, can be photo-switchable, photo-convertible and contribute a wide variety of colors useful for multilabelling and pulse-chase experiments (Adam, 2014).
Both transient expression and genome editing with tools like CRISPR-Cas9 enable the production of chimera proteins containing both the target protein and reporter as a single expressible unit. Epifluorescence techniques can subsequently be employed to track, localize and quantify target proteins with precision. While this labelling method has proven very successful for FM techniques, a robust endogenously expressed inherently electron dense label has yet to be developed/discovered for use in EM (see also later). Generating in vivo expressible electron density, in a manner akin to GFP, for EM visualization has proven to be less straight forward. One of the first methods to generate electron density in this way was the polymerization of 3,3-diaminobenzide (DAB) by a horseradish peroxidase enzyme (HRP). In the presence of H 2 O 2 , HRP catalyzes the polymerization of DAB into a highly insoluble brown precipitate. If the reaction is left to proceed long enough the precipitate can be visualized by standard wide field microscopy. However, for the purposes of EM, the precipitate is not inherently electron dense and therefore does not generate contrast in the electron microscope. The required contrast is generated during sample preparation due to the osmiophilic characteristic of the DAB precipitate. During standard osmification, OsO 4 preferentially binds to the precipitate generating contrast between itself and the surrounding tissue. An early example of this system was labelling the endocytic lumen of l-cells to study pinocytotic flux (Steinman et al., 1974). Since then, a variety of endogenously expressed labels that precipitate DAB have been developed: HRP (Valnes & Brandtzaeg, 1984), APEX (Martell et al., 2012), Mini-SOG (Shu et al., 2011).

| DAB precipitation based probes
HRP contains four disulphide bridges and two calcium binding sites that are essential for its catalytic function. It should be noted that in reducing or low Ca 2+ environments, such as the cytosol, these disulphide bridges do not form and the enzymatic function is largely disabled. Therefore, labelling endogenous cytosolic proteins is not effective with HRP. HRP is also a relatively large protein (44 kDa) for use as a probe (e.g., as compared to 27 kDa for GFP). Typically, larger proteins are disadvantageous as they are more likely to negatively impact the labeled protein's native function and trafficking. On the other hand, the photogeneration of 1 O 2 needed to precipitate DAB, can be introduced without the use of detergent (Kobayashi et al., 2016). There are several genetically-expressed probes that can be converted from a fluorescent to an electron dense signal using photo or chemical conversion (Ando et al., 2018). 2+

| ReAsH and FlAsH and MiniSOG
One of the first such genetically targetable 1 O 2 generators was the biarsenical dye ReAsH. ReAsH (and its counterpart FlAsH) bind with F I G U R E 2 Mapping the fluorescence to the ultrastructure and visualization of HRP-DAB. After the cell of interest has been identified in the TEM ( Figure 1) the ultrastructure can be mapped to correlate with the fluorescence pattern (arrows, a). EGF has been taken up into endosomes where the HRP has been converted into an electron dense precipitate. This precipitate is distributed diffusely throughout the endosomes (arrows, b). Scalebar is 5 μm in (a), 10 μm in (b). high affinity and specificity to an endogenously expressed short tetracysteine sequence tagged to a protein of interest. ReAsH will then fluoresce red and FlAsH green upon binding (Gaietta et al., 2002). To form the targeted localized electron density, photo-conversion is required using intense light in the presence of oxygen. The fluorescent label under these conditions catalyzes the process of singlet oxygen generation which in turn precipitates DAB ( Figure 3).
Unfortunately, the arsenic containing molecules induce cytotoxicity unless treated with antidotes such as 1,2 ethanedithiol (Gaietta et al., 2002). Subsequent papers highlighted these and others issue of non-specific interactions producing high background signals (Machleidt et al., 2007, Hoffman-Kim et al., 2007. The successor of ReAsH called mini singlet oxygen generator (MiniSOG), used a similar approach, using an inherently fluorescent probe that can be enzymatically activated by exposure to blue light, photogenerating 1 O 2 for DAB precipitation (Martell et al., 2017). Min-iSOG is a small protein tag, 106 amino acids in length engineered from Arabidopsis phototropin II. Importantly, MiniSOG is capable of generating fluorescence and singlet oxygen species without the addition of any exogenous cofactors. However, to convert the singlet oxygen species into an electron dense signal, similar to HRP, requires the addition of DAB as a substrate (Figures 2 and 3). Thus far, MiniSOG is the closest example of an entirely endogenous CLEM probe capable of generating fluorescence and electron density. MiniSOG has yet to be widely adopted due to its relatively low quantum yield of 0.37, which is out competed by GFP/EGFP (Shu et al., 2011).

| Ascorbate peroxidase
To address some of the above shortcomings, a novel enzyme was developed based on a class I cytosolic ascorbate peroxidase (APEX).
APEX, a plant peroxidase, supersedes MiniSOG as a better SOG but lacks intrinsic fluorescence and therefore, requiring fusion to a suitable FP for use in CLEM studies (Ando et al., 2018). APEX is active in reducing environments as it does not contain any disulphide bridges so enzymatic function is preserved in the cytosol. It is also much smaller in size (28 kDa) than HRP and similar to the widely used GFP variants. Initially the active site of APEX was unable to accommodate DAB as a substrate so mutagenic work was undertaken to increase the promiscuity of the active site for aromatic compounds. Further mutagenic work needed to be performed to inhibit the formation of APEX-APEX homodimers which would otherwise impinge on the native trafficking/function of any tagged protein (Martell et al., 2012(Martell et al., , 2017. The method of combining two genetic tags together; one fluorescent and one convertible to an electron dense product is a technique previously established with HRP developed into FLIPPER (De Beer et al., 2018;Kuipers et al., 2015), see also later. However, HRP is greatly limited by its requirement for tetramerization, glycosylation and high Ca 2+ concentration, in addition to its poor resolution DAB reaction product that tends to diffuse away from the target site (Hopkins et al., 2000) (see also Figure 2).

| APEX-gold
APEX-gold is a relatively new probing method that follows the simple strategy of converting APEX DAB reaction product into silver/gold particles to increase the signal to noise ratio for EM detection (Rae et al., 2021). In this way, APEX-Gold utilizes the robust, well-defined properties of APEX as a genetic tag complemented by quantifiable particulate gold staining similar to that previously established with immunogold labelling. APEX fusion constructs follow a standard workflow, where following fixation using aldehydes, they are stained with easily diffusible H 2 O 2 and DAB, which is subsequently converted by APEX into an insoluble polymer. The DAB polymer eventually becomes visible upon treatment with OsO 4 (Martell et al., 2017). product less easily distinguishable against the electron dense cellular background thus, APEX-Gold is a suitable solution to this problem (Rae et al., 2021). The Parton group have successfully managed to utilize the standard argyrophilic DAB reaction product to locally convert metal salts; silver nitrate to sub-microscopic metallic silver followed by gold chloride substitution (gold toning), to form 10-15 nm sized colloidal gold particles (Figure 4). They went on to show the compatibility and sensitivity of APEX-Gold with many microscopic techniques including detection of POIs localized to the membrane, cytoplasm and cytoskeleton and interestingly, were able to ascertain proteins expressed close to or below endogenous levels.
This idea has been further developed to test pairwise proteinprotein interactions in fixed and living cells by genetically encoding half of a "split" fluorescent protein to each interacting candidate.
APEX-Gold can be complemented by another technique called Nanobody-APEX; where a GFP-nanobody fused to APEX is cotransfected with a GFP-tagged protein, enabling specific APEX recruitment to GFP-tagged POIs (Ariotti et al., 2015(Ariotti et al., , 2017Rae et al., 2021). Additionally, any unbound APEX construct is sufficiently degraded by the proteasome, thus ensuring a localized source of fluorescence and electron density to the target for CLEM studies without the need to balance expression levels between APEX-nanobody and GFP-labeled constructs (Ariotti et al., 2015(Ariotti et al., , 2017.In this way, proteins exhibiting sufficient proximity will come together to reconstitute the full fluorescent protein and enable emission of photons elicited by the appropriate excitation wavelength (Ariotti et al., 2015(Ariotti et al., , 2017.   (Mercogliano & DeRosier, 2007) and in yeast (Morphew et al., 2015).
Other demonstrations of the MT probe used AuCl, which formed electron dense particles at the target location and was corroborated by subsequent immunolabeling ). However, AuCl is not soluble in the aqueous media which the cells were incubated in (Gammons et al., 1997;Lu et al., 2008). This raises a number of questions with regard to the particles formed and whether the probe is working as intended (Sharp, 2014).

| AFFINITY BASED LABELLING
From the above, it is clear that the currently available probes using endogenous electron density generation are not without issue. Furthermore, none of the endogenous probes can generate electron density with as much contrast and precision, with respect to the actual label, as gold nanoparticles.

| Fluorogold
Unlike endogenously expressed tags which are inescapably linked to the protein of interest, exogenous probes require a mechanism to reach and co-localize with the target. Affinity based probes rely on specific complementary ligand-target binding to co-localize the probe to a protein/object of interest. When labelling endogenous proteins, the cell membrane poses a significant barrier of entry for exogenous F I G U R E 4 mechanism of action of APEX-Gold. The peroxidase activity of APEX, expressed as part of a POI, converts soluble DAB into an insoluble precipitate. Using subsequent silver and gold precipitation, electron dense particles are formed (based on 32-34).
probes. Most commonly cells are permeabilised in some way to disrupt the cell membrane and allow the probe entry. This process can cause ultrastructural damage and cellular content extraction. AuNPs are the most commonly used and well-established label to mark proteins in TEM (Robinson et al., 1998). This is due to the superior contrast they produce in the TEM compared with other less dense metals, their low cytotoxicity (Connor et al., 2005) and their ease of synthesis and surface functionalization. One of the first and most widely cited examples was by Robinson and Takizawa who showed the efficacy of 1.4 nm gold nanoparticles conjugated to antibodies termed nanogold (NG) as a TEM probe (Powell et al., 1998;Robinson et al., 2000). NG also displayed better sample penetration and labelling efficiency compared to colloidal gold due to its smaller size (Robinson et al., 2000). Additionally, antibody fragments can be functionalized with NG without the need for macromolecule stabilization which confers high diffusivity (Takizawa et al., 2015). Fluoronanogold (FNG) combines its predecessor, NG, with an organic fluorophore attached to a region removed from both the target binding site and each other to form a single probe for use in CLEM workflows (Takizawa et al., 2015). FNG labelling on ultrathin cryosections was, to our knowledge, the first example of the exact same structures imaged with both FM and TEM (Takizawa et al., 2015).

| Quantum dots
Quantum dots (QDs) are nanoparticles composed of semiconducting materials most commonly cadmium and selenium. QDs display sizetunable fluorescent emission, resistance to photobleaching and high quantum yields. These inherently fluorescent and electron dense nanoparticles displays very desirable attributes for use as a CLEM probe. Giepmans et al. (2005) explored the feasibility of using QDs as CLEM probe with moderate success. The authors demonstrated triple labelling resolvable in both FM and EM using solely QDs. Despite appearing to be an efficient and effective label there were several key drawbacks. Cells labeled with QDs could not be fixed with either paraformaldehyde/glutaraldehyde or osmium tetroxide. Instead, the cells had to be dehydrated and triple stained with UA after being F I G U R E 5 mechanism of action of APEX-Nanobody labelling. A GFP-expressing POI is labeled using a nanobody coupled to APEX. The peroxidase activity of APEX, expressed as part of a POI, converts soluble DAB into an insoluble precipitate. Using subsequent silver and gold precipitation, electron dense particles are formed (based on 32-34).
F I G U R E 6 mechanism of action of FLIPPER. In the FLIPPER system the POI is expressed with both a fluorescent protein (FP) and an HRP/APEX module. The peroxidase activity is again used to generate an insoluble DAB reaction product (based on 29).  . Cell surface receptors are one ubiquitous example of ligand binding inducing multiple proteins to move proximal to one another triggering a transduction pathway. EphA3 is a cell surface receptor with tyrosine kinase activity found in neuronal synapses. Using quantum dots to label this protein has been shown to activate this protein's transduction pathway. Howarth et al used monovalent quantum dots with a single binding site for EphA3 and was able to avoid proximity induced phosphorylation and pathway activation (Howarth et al., 2008).

| ALTERNATIVE STRATEGIES
Despite revolutionizing microscopy, most fluorescent proteins are hindered by OsO 4 treatment for EM preparation as their fluorescence capability is decreased and photoconversion adversely affected (Kukulski et al., 2011). However, FP variants such as, mEos4 have been demonstrated to withstand fixation and OsO 4 treatment to allow SR-PALM visualization and subsequently in TEM/SEM (Paez-Segala et al., 2015) but retention of the fluorescent signal is still low compared to a non-Osmium treated sample. In principle, such single modality genetic probes would be ideal CLEM probes as they are small, well localized to the target and work in both techniques but are difficult to find. Considering the drawbacks associated with a lot of the approaches described above, people have sought alternative strategies to overcome or by-pass some of those issues. These methods seek to remove the requirement of a specific electron dense signal entirely, favoring the use of the existing fluorescent label in some way to find areas of interest in the EM. With the correct preparation it is possible to retain a fluorescent signal within plastic embedded sections (Nixon et al., 2009). Samples where the fluorescence has been retained inside the resin have usually been high pressure frozen and subsequently embedded in Lowicryl HM20 (Nixon et al., 2009, Kukulski et al., 2011, Lee et al., 2018. Especially the method developed by Kukulski and co-workers has attracted significant attention as it can reach a correlation precision of around 50 nm. This allows for accurate projection of the fluorescent signal on top of the EM ultrastructure and unequivocal assignment of the fluorescent signal to an organelle.

| CONCLUSION
Over the last 20 years, several discoveries have launched both FM and EM techniques to the fore. The development of GFP/fluorescent labels and super resolution techniques have resulted in FM becoming a mainstay in the Biosciences. Furthermore, the resolution revolution (Kühlbrandt, 2014) and advent of direct electron detectors has enabled EM to resolve structures previously only discernible with Xray techniques. There is a draw to develop probes and visualization methods to combine these two approaches, but their development has lagged slightly behind. Currently a combination of avidity based and endogenously expressed probes are used for CLEM workflows.
While antibody-based probes including FNG (Takizawa et al., 2015) have been successfully used to localize antigens in EM for decades, they do present major drawbacks. Firstly, they necessitate the use of cell permeabilization if labelling endogenous targets which can cause various artifacts. Secondly, their large size limits sample penetration and access to their intended targets. Finally, for specific protein targets, antibody avidity can be low resulting in low labelling efficiencies.
Endogenously expressed protein-based labels do not suffer from the above drawbacks, however they bring their own disadvantages. FP constructs can only be used to label other proteins and are unable to discern post-translational modifications. Furthermore, the creation of a construct can have deleterious effects on native protein function and trafficking. At the current time, there are a large enough selection of available probes in order to select one that is optimal for a specific sample. We have summarized the specific pros and cons of each of the discussed techniques in Table 1. However, with the CLEM field constantly evolving new probes are sure to be developed that may supersede the current ones.
It also seems likely that additional techniques may also be added to the microscopist's toolbox in the future (Walter et al., 2021). As technology progresses, mass spectroscopy imaging (MSI) may become a useful and relevant technique. Techniques such as EELS and EDS are able to probe light and heavy elemental composition of a sample respectively. However, MSI would be able to identify any ionisable molecule allowing important biomolecules such as proteins and lipids to be readily identifiable. Furthermore, specific chemical modifications to these molecules, such as post-translational modifications could also be detected. At the current time however, the maximal lateral resolution of MSI is around 1.4 um (Kompauer et al., 2017) which is insufficient to gather meaningful data from most biological samples. This limitation arises from a trade-off between spatial resolution and sensitivity. As the laser diameter decreases (MALDI), so does the number of metabolites ionized producing a weaker signal. If this technology develops to enable a tenfold increase in resolution, one could begin to analyze subcellular structures such as intracellular vesicles, which could see MSI becoming a valuable and widely used methodology.
T A B L E 1 suitability of probes for CLEM workflows

ACKNOWLEDGMENTS
HT was supported by an EPSRC PhD studentship as part of the Bristol Centre for Functional Nanomaterials. We thank the excellent staff of the Wolfson Bioimaging Facility for their continuing support.

DATA AVAILABILITY STATEMENT
Data sharing is not applicable to this article as no new data were created or analyzed in this study.