To understand an electrical circuit, one needs to know which circuit elements are present and how they are organized. Because circuit function is critically dependent upon the arrangement of the elements, it is impossible to understand the function of a circuit without knowing what elements are connected and how those elements are wired. Likewise, to understand the neural circuits of the brain, one must know the activity patterns of a population of neurons as well as the connectivity among those neurons. If the identities of individual neurons can be clarified reproducibly in different experiments, information derived from various studies assessing functions and connectivity can be combined. However, despite recent progress in the systematic profiling of the morphological, physiological and molecular characteristics of neurons, reproducible identification across individual mammalian brains is far from being achieved in full (Peng et al., 2021; Zeng & Sanes, 2017). Therefore, to gain a comprehensive understanding of the neural circuit, an appropriate set of information must be accessed within a subject’s brain. In this short review, we will discuss recent advances in anatomical methods that can be used to aggregate physiological information and present future perspectives on related work.

Electron microscopy (EM) provides sufficient resolution to resolve individual synapses and trace thin neurites. With the aid of rapid developments in artificial-intelligence-based morphological analysis (Beier et al., 2017; Januszewski et al., 2018) and various ingenious high-throughput EM image acquisition technologies, it has become possible to probe individual synaptic connections in cubic-millimeter-scale neural circuits (Table 1). Insightful review articles addressing the detailed advantages and limitations of these high-throughput EM technologies are available elsewhere (Briggman and Bock, 2012; Briggman and Denk, 2006; Helmstaedter et al., 2008; Kubota et al., 2018). Thus, in this minireview, we wish to highlight the technical hurdles facing EM-based circuit reconstruction in regard to retaining physiological information.

Connectomic information on invertebrates such as Caenorhabditis elegans and fruit flies can be understood through comparisons of various functional studies performed in different individuals because the neurons of these organisms are relatively well classified morphologically and genetically (Hobert et al., 2016; Jenett et al., 2012; Yemini et al., 2021). On the other hand, in mammalian brains, neither location nor neurite morphology nor genetic profiling can be used to predict the identity of neurons. Therefore, to understand the circuit elements and connectivity of the mammalian brain, EM images must be correlated with images that show the physiological identity of neurons from the same brain tissue. These physiological images can vary depending on the biological questions to be addressed; the images may depict the firing properties of the neurons during a particular behavior as acquired from 2-photon Ca²⁺ images (Table 1), the long-range connectivity of the neurons from neurotracers, or cell types identified based on the expression of marker proteins (Adams et al., 2016). In most cases, identifying neurons in terms of their functional or molecular properties is a task best suited for fluorescence light microscopy (FM). Indeed, while large-scale EM studies on invertebrates provide rich functional and structural information, those concerning mammalian brains tend to focus on either analyzing structure and connectivity rules per se or adding functional information by means of correlative FM images (Table 1).

Correlative light and electron microscopy (CLEM) is technically challenging for many reasons. In addition to the difficulty of simultaneously preserving ultrastructure and fluorescence, the imaging axes of functional 2-photon microscopy and EM are usually perpendicular to each other (however, see [Andermann et al., 2013]). Thus, the number of neurons whose functions have been measured in the reconstructed volume is inevitably small. Furthermore, because the accuracy of the alignment between EM and FM images is at the level of somata, not the level of neurites, all neurites of the neuron of interest must be reconstructed to examine their connectivity. The results of axon tracing based on anisotropic EM images are often ambiguous because axons are thinner and ~7 times more numerous in a given volume than dendrites (Fiala and Harris, 1999; Kasthuri et al., 2015; Shepherd and Harris, 1998). Even artificial-intelligence-aided dense reconstruction of axons yields what are known as “edge” cases and requires extensive human correction of the segmentation results (Dorkenwald et al., 2019; Kasthuri et al., 2015; Motta et al., 2019).

We believe that the emergence of genetically encoded and EM-compatible tracers has opened a new path for efficiently studying long-range connectivity, molecular identity, and even firing properties with a limited temporal resolution (Atasoy et al., 2014; Han et al., 2012; Li et al., 2010; Lin et al., 2016; Schikorski et al., 2007; Zhang et al., 2019). In particular, genetically encoded dimerized enhanced soybean ascorbate peroxidase-2 (dAPEX2) directed to various subcellular organelles enriched in synapses immediately permits the anterograde labeling of multiple synaptic inputs. Ginty and colleagues showed that transduced dAPEX2 in the mitochondrial matrix of layer (L) 5 neurons leads to distinguishable labeling of the mitochondria throughout cell bodies and neurites without compromising ultrastructural preservation (Zhang et al., 2019). These authors also provided dAPEX2 constructs designed to target various subcellular compartments, including synaptic vesicles (SVs), plasma membranes (PMs), and the endoplasmic reticulum (ER). When the relative subcellular distribution of synapses from multiple brain regions is of interest, as in (Kim et al., 2021b), postsynaptically targeted neurons may be labeled with PM-targeted dAPEX2, while presynaptic groups of neurons are labeled with SV- and mitochondrial-matrix-targeted dAPEX2. Because the synaptic boutons are filled with SV and enriched with mitochondria, the origins of the boutons can be visualized without long and laborious axon tracing. The cell types of neurons, as defined based on the expression of marker proteins, can be visualized in EM by fusing the marker with dAPEX2, as demonstrated in (Martell et al., 2017; Rae et al., 2021). Finally, we suggest that photoactivable calcium-sensitive expression of dAPEX2 can enable selective EM-compatible labeling of neurons that are involved in a specific behavior (Lee et al., 2017). Let us assume that the question to be addressed is the local connectivity among neurons sharing orientation selectivity in the primary visual cortex (V1), as in (Ko et al., 2011). Neuronal mitochondria with specific orientation selectivity can be labeled with dAPEX2 by means of selective visual stimuli coupled with in vivo photoactivation. Because mitochondria are densely present in both axonal varicosities and dendrites, reconstructions around synapses with and without dAPEX2 can provide a statistical conclusion to this question (Fig. 1B, insets). Various light-dependent transcriptional regulation tools (de Mena et al., 2018) with different wavelengths allow orthogonal labeling of multiple activities (e.g., vertical vs horizontal orientation selectivity, see Fig. 1A).

Nonetheless, this method, like others, comes with limitations. Full structural reconstruction may be necessary because most neurons could have an intermediate receptive field and thus express both markers at a certain ratio. Many of the critical determinants of this method, such as sensitivity, half-life, and physiological interference, remain unclear and are difficult to quantify. The lack of fine temporal resolution can be problematic for many experiments because it can obscure essential functional information, such as the pattern of activity and epoch sensitivity of the neurons.

Abbe’s rule and conventional microscopy

The theory of image formation was proposed by Ernst Abbe. The theory suggests that an image arises from a diffraction process, specifically when incoming light waves from an object are diffracted by an objective lens (Hecht, 2016). Abbe found that a larger aperture—and thus, a higher numerical aperture (NA)—results in a higher resolution, as described in the following simple equation:

The NA also affects the resolution in the z (axial) direction:

Another way of describing the resolution of a microscope is the point spread function (PSF), which corresponds to an image of an infinitesimal object blurred by diffraction. The size of the PSF reflects the resolution of the imaging system in terms of both the theoretical and experimental limits. Usually, the full width at half maximum (FWHM) of the PSF matches Abbe’s rule. In an invariant space system (i.e., an aberration-free system such that the PSF is the same everywhere), one can think of an image as a convolution of the imaged object and the PSF (Hecht, 2016).

Effectively, in the field of neuroscience, two enhanced green fluorescent protein (EGFP)-expressing dendrites separated by a lateral distance of less than approximately 205 nm will appear as a continuous structure. Worse, neuronal structures separated by less than approximately 465 nm of axial space will appear to be in physical contact even under ideal conditions when the objects exist on a single imaging plane.

In practice, however, because most neurons have radially oriented neurites, the objects on the focal plane are often obscured by out-of-focus structures. Therefore, an imaging method that can selectively produce clear images of the focal plane within thick brain tissue is required. This is one of the critical reasons why confocal laser scanning microscopy is such a widely adopted imaging tool in neuroscience. In confocal microscopy, optical sectioning is achieved by illuminating an object with focused laser light and detecting the light through a pinhole that eliminates light originating from out-of-focus points. Despite this advantage, confocal microscopy is not appropriate for studying connectivity due to its insufficient resolution, especially along the axial dimension. Many imaging methods have been developed and adopted for neural circuit reconstruction to circumvent this insufficient resolution while preserving the advantages of FM.

Physical methods to circumvent the resolution problem

To circumvent the limited resolution of FM imposed by the properties of light, some physical solutions have been suggested, namely, expansion microscopy and array tomography (Micheva and Smith, 2007; Wassie et al., 2019). A method of isotropic expansion of biological specimens was initially suggested by Boyden’s research team (Chen et al., 2015; Wassie et al., 2019). They took advantage of the expandability of a polymer network. By cross-linking the network of swellable polyelectrolyte hydrogel in brain tissue, they achieved magnification by a factor of 4.5 or more without noticeable distortion of the tissue while preserving the antigenicity of the tissue (Chen et al., 2015; Truckenbrodt et al., 2018). This method effectively allows an optical resolution comparable to that of superresolution microscopy to be achieved using conventional confocal microscopy. However, the typical challenges of confocal microscopy still hinder large-scale circuit mapping. Compared to other FM imaging methods, confocal microscopy requires a long imaging time due to the scanning of the excitation laser, suffers from increased photobleaching because of the prolonged exposure of the fluorophore to the laser, and has relatively low sensitivity because the pinhole tends to block a substantial fraction of the in-focus signal. Because expanded tissues have a lower fluorophore density per unit volume and thus require slower scanning, the problems of confocal microscopy described above will be more severe with expansion microscopy. Combination with selective plane illumination microscopy (SPIM), in which the optical resolution is provided by an orthogonally applied excitation beam in the form of a thin sheet of light (Huisken et al., 2004), could potentially solve some of these problems (Bürgers et al., 2019).

Array tomography (AT) is another physical solution to overcome Abbe’s rule. Because AT relies on iterative wide-field FM imaging of ultrathin serial sections of brain tissue, the axial resolution of this imaging technique depends solely on the thickness of the tissue, not on Abbe’s rule (Micheva and Smith, 2007). Furthermore, because all the pixels have the same chance of antigen–antibody reactions, this method enables quantitative immunohistochemical analysis. A reasonably high (~80%) synapse detection accuracy has been predicted and demonstrated at an isotropic optical resolution of 200 nm (Rah et al., 2013). Furthermore, the minimal thickness of the sections allows repetitive antibody elution and staining (Micheva and Smith, 2007; Micheva et al., 2010b). In practice, various studies have taken advantage of this technique to successfully unveil the subcellular distribution or molecular diversity of synapses (Bloss et al., 2016; Kim et al., 2021b; Micheva et al., 2010a; 2021; Rah et al., 2013; 2015). Nevertheless, despite the advantages described above, AT also has inherent limitations. Because AT is based on wide-field FM, dense reconstruction of local connectivity is difficult to resolve. Furthermore, the wave nature of light imposes a lateral resolution limit on this technique, and thus, the detection of small synapses can be ambiguous. We believe that local circuits could be resolved via AT in combination with Brainbow (expression of fluorescent proteins) (Cai et al., 2013; Livet et al., 2007; Loulier et al., 2014), although this possibility currently remains unexamined experimentally. Superresolution microscopy providing a resolution as high as a few nm on the matching thickness of the serial sections could overcome the resolution limit of this imaging technique. However, because of the long imaging time and fast bleaching of superresolution microscopy imaging, this task is not as easy as it sounds. Below, we introduce a brief conceptual introduction to several widely used superresolution microscopy techniques, such as photoactivated localization microscopy (PALM), structured illumination microscopy (SIM), and stimulated emission depletion (STED), to discuss the advantages and disadvantages of microscopic techniques as ad possible solution for FM-based microcircuit reconstruction.

Recent progress in superresolution imaging (or nanoscopy)

A wide variety of FM techniques have emerged in an attempt to overcome the diffraction limit (Fig. 2; Sahl et al., 2017). Among these superresolution techniques, scanning-based methods can acquire images directly, similar to a confocal microscope. The methods of STED and reversible saturable/switchable optical linear fluorescence transitions (RESOLFT) directly silence diffraction-limited areas such that the effective PSF is reduced to 40 nm (Vicidomini et al., 2018). PALM and stochastic optical reconstruction microscopy (STORM) use the flickering kinetics of a single molecule, known as stochastic kinetics, to computationally detect the molecule’s position. These methods have the highest resolution of all techniques (Betzig et al., 2006; Sahl et al., 2017). To collect a sufficient number of flickering events, special fluorophores and a large amount of imaging (>10,000 frames per image) are required (Khater et al., 2020). In SIM, the sample is illuminated with several rotated and phase-shifted periodic patterns, and the obtained images contain high-resolution information in a resolvable regime, similar to moiré fringe patterns (Gustafsson, 2000; 2005). Through linear image postprocessing of SIM in Fourier space and the reverse Fourier transformation, it is possible to recover an image with a twofold increase in resolution (~100 nm). This process provides sufficient resolution for synapse imaging while requiring only relatively simple and fast computational processes (Khater et al., 2020). On the sample preparation side, SIM is not restricted by labeling and requires a relatively low dose of light, making live cell imaging easier (Sahl et al., 2017). Finally, by utilizing higher harmonic orders with brighter illumination (saturated SIM), the resolution can be increased even further with a nonlinear reconstruction method, which is computationally complex and heavy (Gustafsson, 2005).

Future perspectives

The abovementioned superresolution techniques provide sufficient optical resolution while maintaining the advantages of FM, such as compatibility with functional imaging, genetic or immunochemistry-based molecular labeling, and relatively straightforward image segmentation due to the high contrast of FM. However, in terms of imaging speed, superresolution microscopy imaging is not significantly advantageous over EM. Although the field of view of superresolution microscopy tends to be wider than that of EM, it is still typically smaller than 50 µm. Furthermore, it takes 3-10 min to acquire a one-tile superresolution image (Huszka and Gijs, 2019; Kim et al., 2021a). Therefore, the long imaging time required to obtain images over a sufficiently large volume to address biological questions is a critical limiting factor. In fact, whereas various high-throughput EM imaging methods have been developed and utilized in neural circuit analysis (Table 1), superresolution microscopy has been mainly developed for studying fine structure or dynamic molecular changes in neuroscience (Ji et al., 2008; Maglione and Sigrist, 2013; Tao et al., 2012). SIM seems to be the best fit for large-scale circuit reconstruction because of the following attributes. SIM provides the minimal necessary optical resolution for reliable identification of synaptic connections, i.e., ~100 nm (Gustafsson, 2000; Kim et al., 2021a; Rah et al., 2013; Wu et al., 2013). To take full advantage of FM, multiple fluorescent channels with conventional fluorophores can be utilized. Compared with other superresolution microscopy techniques, the image acquisition speed is relatively fast because of the relatively large field of view (Fiolka, 2013; Gustafsson, 2005). The shortcomings of SIM, preventing it from being an ideal technique, are excessive imaging time and the inability to image thick tissue sections effectively. Three-dimensional sectioning provides approximately twice the resolution of conventional microscopy, approximately 300 nm, and is applicable only to thin samples such as cell cultures, as opposed to thick and highly diffractive matter such as brain tissue (York et al., 2012). A long image acquisition time for a field is unavoidable because SIM uses a series of sinusoidal illumination patterns.

We have recently developed a new imaging technique called structured illumination microscopy of the putative region of interest with ultrathin sectioning (SIM-PRIUS), which combines SIM with AT (Kim et al., 2021a). To reduce the imaging time, SIM images are selectively acquired from regions with putative synapses of interest as determined based on a set of low-magnification images (Kim et al., 2021a). Furthermore, because SIM-PRIUS relies on ultrathin sectioning of the specimen, the axial resolution is determined by the physical thickness of the tissue. Indeed, when the target structure is as sparse as a particular set of synapses (e.g., synapses sourced from a specific brain region), synapse identification can be achieved with greater accuracy and a 95% reduction in imaging time (Kim et al., 2021a). With the rapid development of artificial-intelligence-dependent object recognition techniques, we believe that this concept can enable expeditious and accurate detection from a relatively large volume of brain tissue.

Despite the initial worries about connectomic research highlighted by Seung and Movshon’s argument, many studies have successfully demonstrated the advances that connectomic studies can provide. It is now widely agreed that brains cannot be understood without knowing how neural circuits are organized. A proper understanding of neural circuits can finally be derived when a set of information about molecules defining the cell types and long-range connectivity, physiological characteristics, and local connectivity is collectively understood. In this respect, we believe that FM is the technique that is best positioned to provide the necessary elements if there is sufficient support from digital image processing, automatic hardware control and feedback systems, and molecular tool development.

This work was supported by grants from the KBRI Research Program (21-BR-01-01, 21-BR-01-04, and 21-BR-03-01), from the DGIST R&D Program (21-IJRP-01), and from the Brain Research Program through the National Research Foundation of Korea (NRF) of the Ministry of Science and ICT (NRF-2017M3C7A1048086 and No. 2017M3A9G8084463).

J.C.R. and J.H.C. designed the article and interpreted the relevant literature. J.C.R. and J.H.C. drafted the article.

The authors have no potential conflicts of interest to disclose.

Representative studies using large-scale EM