M. Lakadamyali, ICFO-Institut de Ciencies Fotoniques, Mediterranean Technology Park, 08860 Castelldefels (Barcelona), Spain. Tel: +(34) 935534191; fax: +(34) 935534000; e-mail: email@example.com
Connectomics is an emerging field that aims to generate and analyse neuronal connectivity data with high-throughput to gain knowledge about the brain. The realization of this goal largely depends on the development of innovative microscopy techniques that can produce large-scale three-dimensional images of brain tissue at nanoscale resolution and automated analysis tools that can extract relevant information from these images. This review surveys the challenges associated with generating the images needed for neuronal connectivity maps and how these challenges may be overcome by novel optical methods that allow sub-diffraction-limit imaging.
Since the late 1800s when Ramon-y-Cajal discovered that neurons are the fundamental building blocks of the brain, the precise way in which these neurons are wired together has remained elusive. Given the close link between structure and function, a comprehensive map of neuronal connections, or the connectome, will likely hold important information about how the brain works, or how it fails to work in neurological, psychiatric and learning disorders. However, except for the connectome of a simple invertebrate system (Caenorhabditis elegans), which took a decade to complete, we do not have a complete wiring diagram of neuronal circuits in the nervous system of any other organism. This gap is partly due to the fact that, until recently, we lacked the suitable imaging tools to visualize the circuitry of the vertebrate brain (Lichtman & Denk, 2011). The difficulty is size: neurons can span very large distances (centimetres or even metres), while the typical diameter of neuronal processes (axons, dendrites) and their synaptic contacts can be very small (nanometres). In addition, these thin processes are densely packed inside the brain. Therefore, the images that we obtain must sample large three-dimensional volumes of brain tissue at nanoscale resolution. In addition, these images must reveal which neuronal processes form synaptic contacts with each other. Finally, the imaging speed must be fast enough to scan through cubic millimetres or even centimetres of brain tissue in a reasonable amount of time. Until recently, this challenge seemed daunting for neuroscience, but with modern methods of imaging and sample preparation, researchers are now starting to address this problem in several different ways.
Many studies involving circuit reconstruction and tracing have been carried out by electron microscopy (EM). EM provides exquisite spatial resolution, which is essential for discriminating thin neuronal processes from one another. In addition, synapses can be readily identified in EM images from their morphology and the presence of synaptic densities. However, in this review I will focus on recently emerging and powerful optical methods that also hold great promise for neuronal connectivity mapping. Fluorescence microscopy offers complementary capabilities to EM: (1) high molecular specificity, (2) multicolour imaging, (3) high-throughput and (4) targeted labelling and imaging of specific cellular and subcellular populations. These features combined with the recent development of super-resolution microscopy, which is closing the resolution gap between fluorescence and electron microscopy, will prove highly useful for circuit reconstruction.
Imaging neuronal connectivity with light microscopy
Transgenic mice that express fluorescent proteins in the membrane and/or cytoplasm of their neuronal axons and dendrites are excellent tools for in vivo and in vitro imaging using fluorescence microscopy (Young & Feng, 2004). Together with state-of-the art imaging techniques such as confocal and two-photon microscopy, these mice have provided invaluable information about neuronal function at the cellular, subcellular and molecular level. In the case of the peripheral nervous system such techniques even provided complete wiring diagrams of neuromuscular connections in small muscles (Lu et al., 2009; Fig. 1A). However, until recently, circuit reconstruction in the central nervous system (CNS) using optical methods relied on sparse labelling of neurons in the brain, since the diffraction limited resolution of fluorescence microscopy (∼300 nm lateral and >500 nm axial resolution) makes it difficult to discriminate thin processes in a densely labelled neuropil of the brain. Sparse labelling means that a large fraction of the synaptic partners within a circuit may not be labelled and therefore may not be identified, leading to incomplete circuit reconstruction. Development of novel labelling schemes and sample preparation methods compatible with fluorescence (and electron) microscopy holds the promise to overcome this limitation.
One way to distinguish densely packed neuronal processes is to code each one with a unique colour. If all neurons within a circuit are labelled with a different colour and this colour remains constant inside the entire neuron, branches belonging to the same neuron can be easily identified without the need for long distance tracing. Multicolour labelling of neurons with a remarkable number of distinct colour hues (up to 100) has been realized in the ‘Brainbow’ transgenic mice (Livet et al., 2007; Fig. 1B). The colour coding was achieved by stochastically expressing a combination of three distinct fluorescent proteins in the cytoplasm and/or membrane of neurons. However, given the large number of neurons in the brain, even 100 colours is not enough to uniquely colour-code all neurons within a circuit. Therefore, ‘Brainbow’ mice do not completely avoid the need for tracing neuronal processes from their starting point at the cell soma to their synapses over long distances. When branches belonging to different neurons intertwine with one another, high spatial resolution becomes necessary to disambiguate them. Furthermore, while in some cases the synaptic contacts may be identified by their morphology, additional staining with pre- and postsynaptic markers is essential for identifying all synaptic contacts. Finally, due to the three dimensional architecture of brain circuits, imaging or sample preparation methods that allow reconstruction of large volumes are needed.
One recent method that can achieve 3D volumetric reconstruction is Array Tomography (Micheva & Smith, 2007). This technique takes advantage of automated serial sectioning of plastic-embedded brain tissue into ultrathin slices. The ribbon of serial sections is mounted onto a coverglass, immuno-stained with antibodies and imaged sequentially by fluorescence and/or electron microscopy. 3D volumetric images can then be reconstructed by aligning the images from each serial section. By stripping off the original antibodies and re-applying new ones, a large number of molecules can be imaged and aligned. This technique can in principle be combined with Brainbow transgenic mice to generate 3D volumetric images of both neuronal morphology in multiple colours and the molecular content of synaptic contacts.
In order to generate dense and reliable circuit maps, labelling methods such as ‘Brainbow’ and sample preparation methods such as ‘Array Tomography’ must be combined with novel optical tools that offer high spatial resolution. In the past decade, several techniques have been developed to surpass the diffraction limit of resolution in far-field fluorescence microscopy. These techniques have already found several applications in the field of neuroscience and are starting to be used for imaging and tracing neuronal morphology at high resolution.
Super-resolution fluorescence microscopy
Methods that break the diffraction limit can be grouped into two general categories: (1) those that use spatially patterned illumination, such as Stimulated Emission Depletion Microscopy (STED; Klar et al., 2000), Reversibly Saturable Optically Linear Fluorescence Transition Microscopy (RESOLFT; Hofmann et al., 2005), and Saturable Structured Illumination Microscopy (SSIM; Gustafsson, 2005), and (2) those that rely on stochastic photoactivation, detection and precise localization of individual photoswitchable, photoactivatable or photoconvertible molecules (fluorophores or fluorescent proteins), such as Stochastic Optical Reconstruction Microscopy (STORM; Rust et al., 2006) and (Fluorescence) Photoactivation Localization Microscopy (PALM and fPALM; Betzig et al., 2006; Hess et al., 2006). Below I provide a short description of the two different classes of super-resolution techniques. For more details the reader is directed to recent reviews on this topic (Hell, 2007; Huang et al., 2010).
STED and RESOLFT
In STED, fluorophores in a diffraction-limited volume are excited with a focused beam of excitation light. Since the excited lifetime of a fluorophore is on the order of a few nanoseconds, a second depletion beam (STED beam) arriving within this time can suppress the fluorescence by forcing the fluorophores back to the ground state through stimulated emission. The depletion beam is shaped to resemble a doughnut and aligned with the excitation beam to suppress fluorescence from only those fluorophores that are not in the centre of the doughnut. Therefore, the overall effect of the STED-beam is to shrink the excitation area to a subdiffraction region where the intensity of the doughnut is zero. Stimulated emission is only one of many saturable optical transitions that can be used to send the fluorophore to the dark state and this concept has been extended to other optical transitions through RESOLFT. STED and RESOLFT have achieved lateral resolution as small as ∼30 nm in biological samples (Schmidt et al., 2009; Brakemann et al., 2011; Grotjohann et al., 2011).
STORM, PALM and fPALM
These techniques are based on the idea that the position of a single fluorophore can be determined with nanometre precision by finding the centre of its image (Thompson et al., 2002; Yildiz et al., 2003). In order to overcome the difficulty of localizing the overlapping images of many molecules in a densely labelled sample, the structure of interest is labelled with photoswitchable fluorescent molecules. The fluorescence emissions of the fluorophores can be separated in time by switching on only a sparse subset of them such that their images are not overlapping and therefore their positions can be determined. These fluorophores are subsequently switched off, and repeated cycles of activation and de-activation are used to map the positions of all the fluorophores and reconstruct an image from these positions (Fig. 2A). A wide selection of photoswitchable probes are available in multiple colours that range from organic fluorophores to fluorescent proteins (Fernandez-Suarez & Ting, 2008; Lippincott-Schwartz & Patterson, 2009; Dempsey et al., 2011; Patterson, 2011). These probes have been used to image biological samples with a lateral resolution down to 9 nm (Xu et al., 2011). 3D imaging can also be achieved by determining a molecule's position in all three dimensions. For example, the first demonstration of 3D super-resolution imaging used a simple astigmatic imaging approach. A cylindrical lens is introduced in the image path such that the point-spread function (PSF) of a fluorophore becomes elliptical, and the ellipticity (width of the PSF in x and y) can then be used to calculate the fluorophore's z-position (Huang et al., 2008). Axial resolution down to ∼20 nm has been achieved using this approach (Xu et al., 2011). Alternatively, an interferometric approach (iPALM) using two opposing objectives has achieved a remarkable z-resolution of ∼10 nm, although the imaging depth is more limited in this case (Shtengel et al., 2009).
Super-resolution imaging of neurons and their connectivity
Neuroscience has been among the first fields in which super-resolution imaging techniques have found important applications. STED has been used to image the morphology of dendritic spines at high resolution to determine subtle changes in their structure during synaptic plasticity (Nagerl et al., 2008). Dendritic spines are among the finest structures of neurons, and the neck region of a spine can have diameters as small as 50 nm. Since dendritic spines form functional synapses with axons, the ability to resolve their structure and how it changes over time with high resolution has important implications for many questions in neuroscience, including circuit reconstruction. However, STED has not yet been applied to tracing and reconstructing the connectivity of multiple neuronal partners within a circuit.
STORM has recently been used to trace neuronal connectivity of cultured hippocampal neurons (Lakadamyali et al., 2012; Figs 2B–D). The cultured neurons were labelled by expressing fluorescent proteins as in the case of transgenic mice. The photoswitchable probes were attached to the neurons via labelled antibodies against these fluorescent proteins. An automated approach was used to image multiple overlapping regions in xy and z, and an image registration algorithm based on cross-correlation was developed to align these images into large, three dimensional mosaics (Fig. 2B). Overall, this approach allowed imaging neuronal morphology with ∼40 nm 2D and ∼100 nm 3D resolution. The image resolution was determined by taking into account two important factors: (1) the localization precision of individual fluorophores and (2) the density of label. The localization precision depends on the number of photons collected from single fluorophores and is typically determined by the brightness of the photoswitchable fluorophores, the numerical aperture of the objective, and other factors that affect the light collection efficiency of the optical microscope. The label density sets a limitation on the resolution according to the Nyquist sampling theorem (resolution is determined by twice the average distance between neighbouring localizations; Shroff et al., 2008). In the STORM images of these cultured neurons, the resolution is currently limited by the label density and the quality of the antibodies rather than the intrinsic localization precision of the fluorescent probes used. Targeting the fluorescent protein to the membrane rather than to the cytoplasm gave the highest label density in small neuronal processes with high surface area to volume ratios. With the currently achieved label density and resolution, the tracing was improved by several fold in the cultured neuron system (Fig. 2C). This improvement meant that more intertwined neuronal processes could be spatially separated and identified as distinct when compared to confocal imaging. Combining STORM with multicolour labelling (similar to Brainbow) further improved the tracing accuracy (Fig. 2D).
Since the morphology of neurons alone may not be enough to establish synaptic partners, additional methods to determine pre- and postsynaptic contacts must be combined with the high resolution imaging of the neuronal morphology. Recently, Dani et al. demonstrated that synapses can be identified in brain tissue with high resolution by imaging well-known pre- and postsynaptic markers (e.g. Bassoon and Homer) with STORM (Dani et al., 2010; Fig. 2E). The STORM imaging was extended to a large number of pre- and postsynaptic molecules to determine their organization and variation among different synapses with high precision. In the future, a combination of this synaptic imaging approach with morphological imaging in brain tissue will provide a unique tool for mapping neuronal connectivity of brain circuitry.
Future perspectives and limitations
What does the future of this field hold? It is conceivable that the above-described STORM methods can also be used to trace neuronal connectivity inside brain tissues. Recently, Array tomography was combined with stochastic switching in a proof of concept study to serially image consecutive sections of thinly sliced brain tissue at high resolution (Nanguneri et al., 2012). Similar to the cultured neuron reconstructions, membrane bound GFP and immunofluorescence was used to image the calyx of Held, a model synapse, within a 50 μm × 50 μm × 2.5 μm volume. However, dense reconstructions in brain tissue require higher spatial resolution compared to the cultured neurons due to the higher density of neuronal processes in the brain. The label density achieved in this study is likely not sufficiently high for complete tracing in dense brain tissue. Therefore, it may be necessary to develop new transgenic mice, in particular those where the fluorescent proteins or other tags are targeted to the membrane with high density. In addition, higher affinity antibodies may also need to be generated. When label density is no longer the limiting factor, the image resolution approaches the intrinsic localization precision of the probes used and resolution as high as 9 nm in x–y has been achieved (Xu et al., 2011). It is possible that even higher resolution may be needed for identifying the thinnest processes in dense reconstructions. Therefore, it may be necessary to develop brighter photoswitchable fluorophores that allow higher precision localization. At some point, the size of the label itself becomes the limitation for resolution. For example, antibodies are ∼10 nm and fluorescent proteins are ∼4 nm, and thus in order to achieve resolutions higher than these numbers, smaller probes are necessary.
Typically, large volume imaging of brain tissues requires serial sectioning, which in turn benefits from embedding of tissue in plastic resins. Therefore, methods that preserve the fluorescence of the fluorophores and the antigenecity of the epitopes after plastic embedding will be important. For example, hydrophilic resins (such as glycol methacrylate) were found to provide the best compromise in maintaining fluorescence, reliable sectioning and tissue ultrastructure (Nanguneri et al., 2012). Another important aspect of circuit reconstruction is the imaging speed. Reconstructing even the smallest circuit may require imaging hundreds or thousands of sections. Therefore, the imaging speed of super-resolution microscopy must also be improved for dense reconstruction of large volume samples with high throughput.
Super-resolution imaging in the context of connectomics provides molecular specificity and targeted imaging of specifically labelled neurons with high spatial resolution. However, the spatial resolution of EM is still superior to that of super-resolution fluorescence microscopy. Moreover, EM offers contextual information since heavy metal labelling highlights the cell boundary and subcellular organelles. Therefore, it may be useful to combine the unique advantages of the two techniques. Different methods have been developed for correlative fluorescence and electron microscopy and recently these have been applied to correlate information provided by super-resolution imaging (such as STED and PALM) with information provided with scanning electron microscopy (Watanabe et al., 2011). Further development of these correlative techniques to provide means of robust alignment between the images and routine sample preparation methods compatible with both super-resolution fluorescence and electron microscopy will have a big impact in the field of circuit reconstruction.
Nanoscale imaging techniques, including electron and sub-diffraction-limit fluorescence microscopy, nano- and micro-scale serial sectioning methods, and innovative transgenic tools, have made it possible to generate high-resolution images of brain tissue. Our next challenge lies in combining these techniques in automated ways for high-throughput imaging of the neurons inside normal, developing and diseased brain. In addition, we need to develop computer algorithms that can identify individual neurons in these images and reliably trace their branches in an automated way. Algorithms that employ machine learning, in which a computer is trained based on manually traced datasets, look promising for achieving this task (Jain et al., 2010). Once we tackle the challenge of automated imaging and image processing, we will need sophisticated systems biology approaches to tease out relevant and useful information from terabytes of data. Finally, it will be essential to combine functional studies with structural imaging in order to correlate the structure of brain's connections to their function. Such a large-scale project requires strong collaborations among many disciplines. Physicists, neuroscientists and computer scientists are combining forces to tackle what can be defined as the next big challenge of our century.
The author thanks Prof. Xiaowei Zhuang and Prof. Jeff Lichtman for discussions and critical reading of the manuscript.