Methods for analyzing neuronal structure and activity in Caenorhabditis elegans

Abstract

The model research animal Caenorhabditis elegans has unique properties making it particularly advantageous for studies of the nervous system. The nervous system is composed of a stereotyped complement of neurons connected in a consistent manner. Here, we describe methods for studying nervous system structure and function. The transparency of the animal makes it possible to visualize and identify neurons in living animals with fluorescent probes. These methods have been recently enhanced for the efficient use of neuron-specific reporter genes. Because of its simple structure, for a number of years, C. elegans has been at the forefront of connectomic studies defining synaptic connectivity by electron microscopy. This field is burgeoning with new, more powerful techniques, and recommended up-to-date methods are here described that encourage the possibility of new work in C. elegans. Fluorescent probes for single synapses and synaptic connections have allowed verification of the EM reconstructions and for experimental approaches to synapse formation. Advances in microscopy and in fluorescent reporters sensitive to Ca²⁺ levels have opened the way to observing activity within single neurons across the entire nervous system.

Introduction

In a 1973 paper, Sydney Brenner outlined a program for research on the animal nervous system: “Thus, what has to be done is clear in general outline: i.e., isolatemutants affecting behavior of an animal and see what changes have been produced in the nervous system” (Brenner 1973). Results of the very first experimental forays he and his co-workersmade to test out this idea had already proved its feasibility. After choosing the little nematode C. elegans as a species to study, Brenner found that behavioral mutants were easily obtained. Analysis of nervous system connectivity by serial section electron microscopy (EM) proved to be possible and neurons were traced in one ganglion and in the anterior ventral and dorsal nerve cords in wild-type and in some of the mutants. Defects could be seen in 8 of 15 mutants examined. In one of them, unc-5, the fifth uncoordinated gene defined, it was found that the mutant lacked entirely motor neuron processes in the dorsal nerve cord (Brenner 1973). With this proof of principle in hand, comprehensive programs of behavioral mutant isolation and complete description of connectivity across the entire adult hermaphrodite animal were undertaken and published respectively in 1974 and 1986 (Brenner 1974; White et al. 1986). EM reconstruction of the specialized mating structures of the adult male tail along with some connectivity in that part of the nervous system was published in 1980 (Sulston et al. 1980).

Further progress in this program awaited the development of molecular cloning and sequencing methods so that the functions of the genes identified might be studied. Today we know that many of the behavioral genes Brenner identified encode conserved cellular components that make up the molecular machinery of single nerve cells and synapses, and in many cases, the properties of other cell types as well. Thus Brenner’s program, perhaps unexpectedly, has made an enormous contribution to cell biology. However, the unique properties of the nervous system emerge from its multicellularity, in particular and in great measure, its pattern of synaptic connections. Behavior emerges from the ensemble, resulting from the effects of this connectivity on the electrical and molecular properties of the constituent cells as they dynamically interact with each other. It is to these higher-order functions, unique to the nervous system, that much current research on the nervous system is now focused.

This study is devoted to three methodologies that are uniquely possible in C. elegans: identification of single neurons in living animals, whole-animal connectomics, and comprehensively monitoring neural electrical activity in freely behaving animals. Each of these areas has been facilitated recently by significant technical advances.

The first challenge in understanding the nervous system is identifying the individual neurons. John Sulston discovered that, due to its transparency, the cells, or at least the cell nuclei, in living C. elegans could be seen by taking advantage of the contrast in the specimen produced with Nomarski differential interference contrast (DIC) microscopy (Horvitz and Sulston 1990). He found that the cells were mostly located in stereotyped positions and so could be uniquely identified in this way. Mostly, but not entirely, as there remains a certain amount of variability or ambiguity. Furthermore, especially in the ganglia of the nervous system, many nuclei were crowded together and a steep learning curve to discern the patterns stood between the beginner and the accomplished worker. The first section of this study describes how this difficulty is overcome by selectively labeling cells with various fluorescent proteins (FPs), approaches that have recently been enhanced, especially for identification of neurons.

Complete sets of connections between neurons can still only be determined by EM. After John White’s 1986 publication, connectomics in C. elegans by EM was made more accessible by the development of computer software that allowed scoring electron micrographs (EMs) at the computer screen (Jarrell et al. 2012; Xu et al. 2013). The general field of connectomics by EM across multiple species has now been taken up in many laboratories and technical advances have been introduced to speed the process, which nevertheless is still forbidding. These advances have covered the range of methodologies, from sample preparation, serial sectioning, image acquisition, and image analysis. They are beginning to be applied to C. elegans and other nematodes and are briefly described in the next section with references to published protocols. EM reconstructions are limited to very small sample sizes. This has motivated the development of synaptic labeling methods so that specific synapses may be examined in multiple animals and genetic investigations of synapse formation can be carried out. These methods, which have provided an important verification of the EM reconstructions, are described next.

The connectomes that have been generated now for several organisms along with C. elegans are complex networks of connections. This has brought about the necessity of employing the methods of graph theory, the branch of mathematics that deals with the properties of networks, to analyze them. Simple questions, such as how many neurons does a given neuron make connections with and what is the overall distribution of these values across all the neurons, are covered by the graph-theoretical concepts of node degree and degree distribution. In the next section, methods for analyzing the EM connectomics data are covered, from 3D reconstructions of single neurons to visualization of networks and analysis of their properties by graph theory, along with associated available software packages.

Finally, the connectome is a static structure. But neural activity is a dynamic process. In the final section, recent advances are described that have been made in techniques for visualizing neural activity during behavior. Once again, worm transparency is the key and fluorescent indicators for Ca⁺⁺ concentration, neurotransmitters, and voltages, are available. The development of high-speed image acquisition techniques has reached the exciting stage where the activity of a very large proportion of all the neurons can be assessed simultaneously as a worm moves about its environment motivated by its various needs, including copulation by the male. Accounting for such data will provide the ultimate test for assessing the ability of theoretical models to understand and predict behavior based on nervous system structure and the properties of its constituent cells.

Identifying neurons

A history of neuron identification in C. elegans

A heroic set of experiments that united EM reconstructions with DIC microscopy (also known as Nomarski microscopy) inaugurated neuron identification in C. elegans. The EM studies showed that the worm nervous system consists of a stereotyped set of identifiable neurons, and that they are at roughly reproducible locations in individual animals. These EM images were correlated with the cell nuclei seen in DIC imaging, thus allowing determination of the set of cell lineages that give rise to each neuron. This remarkable finding of neuronal stereotypy in C. elegans sets the worm apart from other model organisms and makes neuron identification possible. Notably, this stereotypy permits researchers to repeat the same experiment, interrogating the exact same neuron(s) in vivo in animal after animal through activity reporters (Lin and Schnitzer 2016), laser microsurgery (Bargmann and Avery 1995), or genetic ablation or activation (Shaham and Horvitz 1996; Chelur and Chalfie 2007; Qi et al. 2012), including by optogenetic (Husson et al. 2013), or chemogenetic (Pokala et al. 2014) approaches, thus opening up the possibility of dissecting the nervous system function at the cellular and circuit levels.

Eileen Southgate, Nichol Thomson, Samuel Ward, and John White reconstructed the neuronal morphology and connectivity of the C. elegans nervous system using overlapping EM sections from multiple animals (Ward et al. 1975; White et al. 1986). Ward’s initial work indicated that the anterior nervous system exhibits identical sets of neurons, identifiable by the gangliar positions of their somas and cellular morphologies, and hinted that this might extend to the remainder of the nervous system. The whole-animal C. elegans EM reconstruction used five animals, with overlapping sections, to provide definitive evidence of this positional and morphological somatic stereotypy. Three animals provided overlapping head sections and two animals provided overlapping tail sections. In illustrating neuronal morphology, the EM reconstruction further provided a complete view of all neurite paths (also known as projections or processes), and in the case of the amphids, their complex dendritic patterns.

In tandem, a series of papers by Sulston and co-workers used DIC microscopy to determine the stereotyped embryonic and postembryonic lineages of all somatic cells within the worm, among these its neurons (Sulston et al. 1975, 1983; Sulston 1976; Sulston and Horvitz 1977). DIC was used to follow all cell divisions leading to the adult hermaphrodite and male. These data were further enriched with the identities of the dopaminergic neurons, using formaldehyde-induced fluorescence (FIF) (Sulston et al. 1975). A substantial set of illustrations accompany the work, providing a map of cell positions, their identifiable characteristics when using DIC, and the locations of the dopaminergic neurons.

Soon after these EM and DIC studies, a variety of techniques emerged to make neuron identification easier and more certain. Subsequent publications used FIF to further identify other members of the monoaminergic neuron classes (Horvitz et al. 1982). Fluorescein (Hedgecock et al. 1985) and lipophilic (Tong and Bürglin 2010) dyes were used to identify a subset of amphid, phasmid, and even inner labial neurons. Concurrent work expanded the toolkit to include antibody staining and lacZ protein fusions (Siddiqui et al. 1989; Fire et al. 1990; Mclntire et al. 1993). These efforts culminated with the introduction of green fluorescent protein (GFP) to label cells by using genetic reporters that drive transgenic GFP expression (Chalfie et al. 1994). FP labeling, combined with DIC images, remains the most popular method of neuron identification today.

Before differentiation, embryonic neuron identities are only accessible by following the divisions that make up their lineage. The advent of affordable personal computers provided a means of recording and following these lineages and thus facilitated embryonic neuron identification. The first software available for this purpose was Biocell which enabled recording and manually tracking embryonic cell divisions using DIC images (Schnabel et al. 1997). Almost a decade later, Starrynite inaugurated automated cell lineaging (Bao et al. 2006). This software used a fluorescent histone label and computer vision to follow embryonic cell divisions up until the 350-cell stage. NucleiTracker4D was later introduced to provide a semi-automated approach that goes beyond the 350-cell stage, tracking embryonic cells through morphogenesis (Giurumescu et al. 2012).

Concurrent with these software advances were widespread efforts to identify gene expression in postembryonic neurons using FP markers. Initially, insufficient markers existed to label and identify the broad variety of C. elegans neurons. This gap was gradually filled by independent research efforts within the worm community, often as a side effect of other research goals. Fluorescent neuronal expression was used to identify neurons by matching their position, morphology, and projection patterns to those in the EM reconstruction, as well as by dye-filling and crosses to other reporter strains, as the toolkit of FPs grew to include other distinguishable fluorophores (Heim et al. 1994; Heim and Tsien 1996; Matz et al. 1999).

Fluorescent reporter strains, usually combined with DIC, provide a straightforward means to identify neurons. Using these strains, neuron identification proceeds by hypothesizing neuron identities based on position (and, when present, cytoplasmic morphology), then crossing the strains to reporter strains to test these hypotheses. WormBase provides a database of reporter strains available for use in neuronal identification (Stein et al. 2001). Many of these strains are accessible from the Caenorhabditis Genetics Center (https://cgc.umn.edu).

Despite these innovations, identifying complex patterns of expression often required a substantial amount of time, involving multiple crosses in order to test hypotheses of neuron identity. Moreover, the collection of neuron identification strains still lacked the ability to identify several neurons in the dense region of the ventral ganglion, and annotations of strain expression often contained ambiguities (e.g., using such phrases as “expressed in other unidentified neurons in the head and tail”), or even errors. To partially alleviate this, three neurotransmitter reporter strains (labeling the cholinergic, GABAergic, and glutamatergic neurons) were circulated, along with carefully verified neuron identification maps for the hermaphrodite and male neurons expressing these reporters (Serrano-Saiz et al. 2013, 2017; Pereira et al. 2015; Gendrel et al. 2016). This reduced the work of neuron identification to, at most, three crosses; but it failed to resolve identification issues in the ventral ganglion, a dense collection of 32 neurons, nearly all of which solely express the cholinergic reporter. Moreover, using these strains still required substantial cell-identification expertise.

Recent innovations by two groups attempt to fill this gap (Toyoshima et al. 2020; Yemini et al. 2021). Their work also addresses a substantial bottleneck in the rapidly advancing field of whole-brain neuronal activity imaging, identifying all neurons within in vivo recording volumes. Their efforts reveal substantial variability in postembryonic neuron positions, thus affirming that the majority of these neurons cannot be unambiguously identified by position alone. Both groups used multicolor reporter strains for neuron identification with semi-automated software. One solution, NeuroPAL (a Neuronal Polychromatic Atlas of Landmarks), disambiguates all neurons in the hermaphrodite and male, at all stages of larval development; identification can be done manually or with accompanying software to help automate the task. Another solution, the JN3039 strain, disambiguates most neurons in the hermaphrodite head but excludes all midbody and tail neurons; identification requires accompanying software. With these advances, multiple computational groups are moving quickly to automate neuron identification using these strains (Bubnis et al. 2019; Nejatbakhsh et al. 2020; Varol et al. 2020; Chaudhary et al. 2021)—thus removing one of the last obstacles in the task of neuron identification: the need for human expertise and oversight.

Methods

Morphology and position:

Nuclear morphology combined with position can be used to identify some neurons. Neurite paths are unique to each neuron, and when visible (i.e., in EM reconstructions, with dyes, or using fluorescent reporters), neurites together with somatic positions can be used to identify all neurons (White et al. 1986). In some cases, the neurites alone are sufficient for neuronal identification; FLP and PVD, for example, display extensive branching in the head and across the whole body, respectively, that uniquely identifies these neurons. The shape of neuronal somas and nuclei may also provide sufficient information for identification, although this has not yet been rigorously explored. Nonetheless, EM and fluorescent data suggest that neurons have stereotyped somatic and nuclear shapes; the SABVs, for example, appear to have smaller nuclei than neighboring neurons in the retrovesicular ganglion (Altun and Hall 2002; Yemini et al. 2021; Figure 1A).

Figure 1.

(A) The head and tail of a NeuroPAL worm. NeuroPAL is a complex, multicopy transgene that expresses combinations of four distinguishable, nuclear-localized fluorophores from 41 neuron-specific promoters. It is integrated at a single chromosomal locus to facilitate use in genetic crosses. In a NeuroPAL worm, every neuron class, and in some cases sub-classes (e.g., distinguishable left/right homologs), expresses a different combination of fluorescent colors, and can be identified by this combination together with positional information (Yemini et al. 2021). In the figure, the four NeuroPAL fluorophores are pseudo-colored to create a red, green, and blue primary color palette. None of the NeuroPAL reporters emit in green—the green reporter in the above images is a pseudo-colored red-emitting fluorophore (CyOFP1). Thus the NeuroPAL fluorophores are distinguishable from GFP, CFP, and YFP, and can identify neuronal expression in reporter strains using these fluorophores or can identify neuronal activity when using GCaMP. As an example of a neuron class with distinguishable nuclear morphology, the SABVs (white arrow) have smaller nuclei than neighboring neurons. (B) DIC images of speckled neuronal and glial cells, compared to those of muscles, intestine, glands, the HMC, excretory cell, and hypodermis. After the L1 larval stage, neurons (red circles and inset) and glia (blue circles and inset) are the only cells that exhibit speckled nuclei (also described as a granular or stippled pattern), termed NUN bodies (Pham et al. 2021). Comparatively, nuclei of muscle, intestine, gland, HMC, excretory, and hypodermis cells, are larger, oval-shaped, and show prominent nucleoli. Scale bars indicate 20 microns in the top row and 1 micron in the bottom row of single nuclei micrographs. ©2021 Pham K, Masoudi N, Leyva-Díaz E, and Hobert O. Originally published in Genetics. https://doi.org/10.1093/genetics/iyaa016. (C) Fluorescent dye-filled neurons. The head and tail of a DiO stained worm. Neurons are labeled with their names.

The nuclei of neurons and glia are distinguishable from those of other cell types in DIC images by their relatively smaller size and flat, speckled appearance (also described as a granular or stippled pattern) (Figure 1B). Recent work has investigated the molecular underpinnings of this speckled appearance, which is due to bodies termed NUN (NUclear Nervous system-specific) bodies, but the nature of these bodies remains a mystery (Pham et al. 2021). Nuclear speckles appear in all embryonic cells; after the L1 larval stage, they remain only in neuronal and glial nuclei. Neurons and glia show similar speckled patterns but can often be distinguished because with the exception of the GLRs, most glia is positioned sufficiently far from neurons.

EM reconstructions of the N2S, N2T, and N2U adult hermaphrodites indicated variability in neuron nuclear and cell body positions, particularly in the retrovesicular ganglion (White et al. 1986; Altun and Hall 2002). Recent work by two groups has quantified this positional variability and found that it can be substantial for many neurons in the head and tail (Toyoshima et al. 2020; Yemini et al. 2021). They provide coordinates and variances for all neuronal nuclei in the head, and Yemini et al. provide these for the tail as well. Those with extensive expertise can recognize neurons by their position in sparse areas such as the pharynx, midbody, and portions of the tail as well as at several gangliar boundaries in the head with positional stereotypy (e.g., the ASK, ADL, and ASI neurons which abut the dorsal boundary of the lateral ganglion—see Figure 1, A and C). Nonetheless, most neurons require additional criteria to confirm their identity. This is especially the case for neurons in the ventral and retrovesicular ganglia, as well as those present at posterior positions in the lateral ganglion, all of which show substantial positional variability.

Embryonic and postembryonic lineaging:

Embryonic and postembryonic lineaging can be used to identify neurons (Sulston 1976; Sulston and Horvitz 1977; Sulston et al. 1983); but both require substantial time and expertise. Several software entries exist to aid in embryonic lineaging. Biocell can be used with DIC recordings and manual annotation to determine cell lineages and thus neuronal identities (Schnabel et al. 1997). Starrynite is used with a fluorescent histone marker and automates the process of cell lineaging, but requires manual review to correct any errors (Bao et al. 2006). NucleiTracker4D also uses a fluorescent histone marker, and using semi-automated methods, it can follow cells beyond the 350-cell stage to determine neuron identities through morphogenesis (Giurumescu et al. 2012). Due to the time and expertise involved in embryonic lineaging, this type of work often uses smaller sample sizes or restricts analysis to shorter developmental time periods. During postembryonic development, fluorescent reporter strains are sufficient to identify neurons, and thus postembryonic lineaging remains primarily useful when investigating cell-fate mutations. Methods that employ a microfluidic chip and image registration can aid in postembryonic lineaging (Keil et al. 2017).

WormAtlas:

Among WormAtlas’s many resources is an index page with information for all C. elegans neurons (https://www.wormatlas.org/neurons/Individual%20Neurons/Neuronframeset.html). This page lists all hermaphrodite and male neurons along with their lineage origin, and provides a brief description and links to neuron-specific pages with further details. The individual neuron pages provide 3-dimensional renderings for the embryonic neuron position, postembryonic position, morphology (often accompanied by representative fluorescent images), and also the dendritic shapes for many of the sensory neurons. These pages and their renderings provide sufficient details to identify any neuron in the worm, as long as its complete morphology is clearly visible.

EM traces:

Tracing EM images provides a 3-dimensional representation of cells. Neurons can then be identified by their position and morphology. The field of EM tracing is rapidly evolving and multiple tools exist to partially automate the task. The Connectomics section provides an in-depth review of the field and available tools. In brief, the presence of long thin processes that extend from the soma (potential neurites) indicates that a cell might be a neuron; although, glia and other cells such as head mesodermal cell (HMC) also display these processes. Thus, some tracing is necessary to identify neurons. Within short sections of the serial EM reconstruction, neurons are identifiable by the presence of varicosities (swollen regions) along with their processes, filled with vesicles, and electron-rich darkly stained presynaptic specializations where the membrane adjoins postsynaptic partners (Figure 2). In C. elegans, chemical synapses are en passant, connecting adjacent neuronal processes (as opposed to being present solely at the terminal ends of neurites). Some neurons are identifiable by further specializations; for example, touch receptor neurons exhibit 15-protofilament crosslinked microtubules in their EM sections, as is noted by the suffix M in their name (e.g., ALM, the Anterior Lateral Microtubule neuron) (Chalfie and Thomson 1982).

Figure 2.

EM image containing neuronal processes. (A) Neurons are recognizable in EM images by varicosities along their processes filled with vesicles (magenta and yellow arrows) and electron-rich darkly stained presynaptic specializations where the membrane adjoins postsynaptic partners (cyan arrow) (White et al. 1986). (B, C) Touch neurons are further uniquely identifiable by their 15-protofilament microtubules (red arrows and insets) (Chalfie and Thomson 1982). ©WormAtlas and ©1982 Chalfie M and Thomson JN. Originally published in Journal of Cell Biology. https://doi.org/10.1083/jcb.93.1.15. Adapted with Martin Chalfie’s permission.

Dye-filling:

WormAtlas provides protocols for dye-filling (https://www.wormatlas.org/EMmethods/DiIDiO.htm). Multiple dyes are available for dye-filling and two are particularly popular, DiO and DiI. These dyes have a peak excitation and emission similar to those of GFP and TagRFP, respectively, and thus can be imaged with the corresponding filter sets. Dye-filling can be used to identify six amphid (ADL, ASH, ASI, ASJ, ASK, and AWB) and two phasmid (PHA and PHB) neuron classes (Figure 1C) (Hedgecock et al. 1985). A minor modification to the DiI staining protocol labels the two inner labial (IL1 and IL2) neuron classes as well (Tong and Bürglin 2010). Dye-filled neurons are identifiable by their position and morphology. ASK, ADL, and ASI are positioned anterior to posterior in order, and they can be further identified by the paths of their neurites and their dendritic morphologies. PHA and PHB can be difficult to distinguish due to their similar position and morphology; but the PHA soma is most often anterior-ventral to that of PHB (Yemini et al. 2021).

Fluorescent reporters:

Neurons whose identity remains unknown but whose cytoplasm or membrane are fluorescently labeled can be identified by their neurites (as long as these are clearly visible and can be disambiguated from other neuron’s neurites) and, in the case of the sensory neurons, via their unique dendritic endings (Figure 1C; Ward et al. 1975). Several nonneuronal cells, such as the GLRs and HMC, also exhibit processes that look like neurites; therefore other morphological information is sometimes necessary to distinguish neurons from these nonneuronal cells.

Neuron-specific reporter strains can be found on WormBase (https://wormbase.org) by searching “for an anatomy term” using the neuron name to access a list of “expression markers,” transgenes that express in and identify this neuron. Separate entries exist for individual neurons and their class (e.g., the ASEL and ASER neurons have their own entries, separate from each other and the ASE neuron class), thus one should search all entries for suitable transgenic strains. These strains can often be ordered from the CGC, and when not available from this central repository, the lab of origin should be contacted.

To aid in the identification of neurons expressing a fluorescent marker in an unknown neuron, a set of neurotransmitter fosmid reporter strains, independently targeting cholinergic, GABAergic, and glutamatergic neurons, have been developed (Serrano-Saiz et al. 2013, 2017; Pereira et al. 2015; Gendrel et al. 2016). The reporter transgenes in these strains were integrated into separate chromosomal locations; these integrants were then crossed to HIM mutations so that the strains generated ∼50% males, thus aiding in crossing them to other strains with unknown expression patterns. OH13470 and OH13646 contain fosmids driven by the promotor of the cho-1 gene to label cholinergic neurons with YFP and mCherry respectively. OH13104 and OH13105 contain fosmids driven by the promoter of the unc-47 gene to label GABAergic neurons with YFP and mCherry respectively. OH12312 and OH13645 contain fosmids driven by the promoter of the eat-4 gene to label glutamatergic neurons with YFP and mCherry respectively. All strains can be ordered from the CGC (https://cgc.umn.edu). Neuron identification is performed as follows: (1) the neurotransmitter identification strains are chosen such that their fluorophore is distinguishable from that of the strain whose expression is being identified (e.g., if the strain with unknown expression uses a GFP reporter, the mCherry neurotransmitter identification strains are used); (2) the strain with unknown expression is crossed to the three neurotransmitter identification strains expressing the distinguishable fluorophore; and (3) the three neurotransmitter expression maps are then used to identify overlapping expression of the unknown reporter strain.

All-in-one identification strains:

Recent innovations by two groups have produced all-in-one C. elegans neuron identification strains (Figure 1A;Toyoshima et al. 2020; Yemini et al. 2021). With these strains, a single cross can be used to identify many (using JN3039) or all (using NeuroPAL) neurons. Both strains can identify neuronal expression using GFP, CFP, and YFP reporters (Cranfill et al. 2016), or neuronal activity using sensors such as GcaMP or Cameleon (Lin and Schnitzer 2016). These all-in-one identification strains can be imaged with nearly any microscope (e.g., widefield, confocal, light-sheet, and structured illumination imaging systems), using commonly available filter sets.

The NeuroPAL strain (available from the CGC) provides manuals and software (both available at https://www.hobertlab.org/neuropal/) to configure any microscope for imaging and to identify all neurons in C. elegans, either manually or with its semi-automated software. NeuroPAL identifies all neurons across all postembryonic developmental stages, and in males as well (unpublished). NeuroPAL can also identify the nonneuronal Amso and Phso1 glia. At present, semi-automated identification only works for the head and tail of adult hermaphrodites, the remainder must be done manually in the software.

The JN3039 strain (available upon request from the Iino and Ishihara Labs) is used with their software ROIEdit3D (available as a supplement to their publication and maintained at https://github.com/YuToyoshima/RoiEdit3D) to identify nearly 60% of the worm’s neurons. Specifically, this strain identifies just over 80% of cells in the adult hermaphrodite head (excluding the retrovesicular ganglion). This strain also identifies the nonneuronal GLR glial, XXX atypical hypodermal, and pharyngeal gland cells, which are not marked by NeuroPAL.

Neuronal fate determination and mutant analysis:

Mutations can affect cell fate, and thus neuron identity. The topic of neuron identity is a contextual one that can encompass lineage, gene expression, morphology, and even behavioral phenotypes. Nevertheless, neuronal fate losses or switches are often evaluated through neuron-specific identity markers. For example, a loss of fate for touch neurons can be evidenced by the loss of their identifying microtubules [e.g., as seen in mec-7(e1506) mutants] (Chalfie et al. 1981; Chalfie and Sulston 1981). As another example, fate switches between the ASEL and ASER neurons is evidenced by left- or right-sided switches in the expression of their identifying genes [e.g., as seen in lsy-2(ot64) mutants which express the ASER-specific marker gcy-5 in ASEL, indicating that its fate has switched] (Johnston et al. 2005).

Most commonly, several identity markers are used to test the hypothesis of losses or switches in fate. NeuroPAL affords a quick means to screen mutants for fate changes and thus generate hypotheses (Yemini et al. 2021). To do so, the mutation is introduced into NeuroPAL (e.g., through a cross) and then all neurons are evaluated to ensure the correct amount are present and that their colors match the wild-type colormap. Both lineaging and NeuroPAL can also be used to hypothesize neuron duplications, by observing an extra division among the stereotyped lineage diagram (Chalfie and Sulston 1981; Chalfie et al. 1981), or by observing an extra neuron that has NeuroPAL coloring and nuclear morphology similar to a nearby neuron. However, both of these methods can benefit from further verification, using other neuron-specific identity markers to ensure a duplication has occurred.

Automated identification software:

Multiple groups have developed software for semi-automated cell and neuron identification. The techniques they use represent a mixture of expertise (e.g., computer vision, statistics, optimization, machine learning, and neural networks). A full discussion of these algorithmic techniques is beyond the scope of this text. Published software incorporates these methods, and thus shields users from any requirement of understanding the internal algorithmic details. Nonetheless, here we provide a brief overview of commonly used techniques for those readers with sufficient background to understand them.

Initially, cell-identification software for the worm focused on identifying nonneuronal cells, such as muscles, which have very stereotypic positioning (Liu et al. 2009b; Long et al. 2009). With the advent of all-in-one neuron identification strains, however, the focus has shifted to neuron identification (Bubnis et al. 2019; Nejatbakhsh et al. 2020; Varol et al. 2020; Chaudhary et al. 2021). The recent crop of new software efforts has only been evaluated on their own individual datasets, thereby precluding a head-to-head comparison of their effectiveness. Still, all of these software algorithms feature most (or all) of the following sequence of steps:

1. Neuronal atlas construction: A neuronal atlas captures the neuron position statistics, their colors (i.e., their fluorescent reporter expression), and potentially, their pairwise positional relationships (e.g., PHA is most often found anterior-ventral to PHB). The atlas is later used to model identities for observed neurons (Bubnis et al. 2019; Varol et al. 2020). Alternatively, a set of atlases can be used to vote on the identities of observed neurons (Toyoshima et al. 2020). Software to create atlases of neuronal positions and colors is available at https://github.com/amin-nejat/StatAtlas/tree/master/Atlas (Bubnis et al. 2019; Varol et al. 2020).

2. Image preprocessing: The image preprocessing step removes noise (e.g., salt-and-pepper noise), usually by smoothing the image (Toyoshima et al. 2016). Preprocessing may then remove nonneuronal cells and features such as gut granules and hypodermal nuclei, either manually or by modeling them as background (Yemini et al. 2021).

3. Neuronal detection: The neuronal detection step of any given algorithm identifies the location of all neurons in the image. Cytoplasmic neuronal morphologies are too dense to distinguish, thus all algorithms to date are designed to work with images of nuclei. These nuclei are still densely packed, and therefore simple computer vision algorithms (e.g., adaptive thresholding and Watershed segmentation) often group several together, requiring further refinement to separate joint nuclei (Long et al. 2009). Thus, many algorithms instead approximate neuronal nuclei as ellipsoids, often using some form of 3-dimensional Gaussian approximation. In these cases, the image is then modeled as a set of ellipses and a variety of algorithms can be used alone or in combination to locate these neuronal ellipses [e.g., ellipsoid fitting (Toyoshima et al. 2016), Gaussian Mixture Models (Toyoshima et al. 2016), and deconvolution (Yemini et al. 2021)]. Once neurons are detected, their colors are measured from their corresponding pixels, and other identifying characteristics, such as their size and shape, may be measured as well (Toyoshima et al. 2016; Nejatbakhsh et al. 2020). Many algorithms exist for detecting nuclei-like blobs in 3-dimensional images, and new methods are being published at a rapid pace, thus a full discussion of these techniques is beyond the scope of this text; nonetheless, a review of several older techniques (including ones that use machine learning, graph cuts, peak detection, and gradient flow) can be found within Toyoshima et al. (2016), and several of the latest deep generative models and neural-network-based techniques are reviewed by Dunn et al. (2019) and Yang et al. (2020).

4. Neuron identification: Neuron identification proceeds by searching for the best match between detected neurons and those represented in the atlas(es), based on identifying statistics. The best match can be encoded as a set of deterministic or probabilistic neuronal identities. Matching individual neurons to an atlas is a linear assignment problem with efficient algorithms to solve it (e.g., using the Hungarian or Sinkhorn expectation-maximization methods) (Bubnis et al. 2019; Nejatbakhsh et al. 2020; Yemini et al. 2021). Matching pairwise neuronal relationships to an atlas is a quadratic assignment problem and, due to its combinatorial nature, requires approximation solutions (e.g., using a Conditional Random Fields model) (Chaudhary et al. 2021). Several popular identification algorithms can further update and improve their probabilistic assignments in response to manual identification by users (Toyoshima et al. 2020; Yemini et al. 2021).

Current limitations of automated neuron identification:

One challenge in evaluating progress in neuron identification has been the absence of a public dataset to compare competing algorithms. The recently published NeuroPAL paper (Yemini et al. 2021) provides a set of 10 heads and 10 tails with labeled neurons (https://zenodo.org/record/3906530), and thus introduces a benchmark for algorithmic comparisons. The NeuroPAL neuronal identification software has 86% accuracy in identifying all head neurons and 94% for all those in the tail; manual annotations improve this accuracy. This software can further identify several neurons at the anterior and posterior of the ventral nerve cord. Nonetheless, to date, no algorithm exists to identify all 74 neurons in the midbody of adult hermaphrodites, between the head and tail. These represent a particular challenge due to the deformable nature of the worm which can take on linear, sinusoidal, circular, omega, and other twisted shapes that substantially curve this midbody region; in contrast, the head and tail are sufficiently rigid across their short expanse so as not to present a problem. Beyond this, no atlases currently exist to identify neurons at other developmental stages and in male worms. This problem should be easily solved by annotating sufficient images of animals representing the missing stages and sex, then using these annotated images to create the missing atlases.

Future methods for neuron identification:

While it is difficult to predict the future of any field, several methods show promise for improving the task of automated neuron identification. Of particular interest are algorithms that do not use color information and thus can be used to augment NeuroPAL, JN3039, or neuronal reporter strains—and perhaps even replace them. Individual neurons exhibit unique characteristics in their GCaMP activity waveform (Figure 3) that may aid in identifying them (Kato et al. 2015), with some limitations when using mutant animals (Kotera et al. 2016; Yemini et al. 2021). For example, sensory neurons (e.g., ASE and AWA) often show high-frequency responses to stimuli, while premotor interneurons (e.g., AVA and AVB) show spontaneous, sustained low-frequency waves of activity. Thus neuronal activity can provide spectral information that could be used, in combination with other information (e.g., position), to aid in identification. These spectral signatures may even provide sufficiently unique “fingerprints” to identify neurons without any additional identification information (i.e., without using color), and thus enable the construction of a “dictionary of activity” for neuron identification. More speculatively, recent algorithms that predict the identity of fly neurons via their position and neurite geometry (e.g., BlastNeuron and NBLAST) (Wan et al. 2015; Costa et al. 2016), and ones that predict the neurotransmitter from EM sections (Eckstein et al. 2020), might be extended in some form that would have predictive utility for worms as well. These new opportunities, along with continuous improvements of existing neuron identification software, provide a promising future for automated neuron identification.

Figure 3.

Neurons show distinguishable signatures of neuronal activity. A 4-minute recording of neuronal activity in a NeuroPAL; GCaMP6s (OH16230) worm (Yemini et al. 2021). The sensory neurons (e.g., ASE and AWA) show high-frequency responses to stimuli, while the premotor interneurons (e.g., AVA and AVB) show sustained low-frequency waves of activity, which are correlated with other interneurons. These network states have been interpreted as spontaneous motor commands for backward- and forward-crawling (Kato et al. 2015).

Connectomics

Determining connectivity by electron microscopy

The basic techniques that were developed in the 1950’s for examining biological ultrastructure with the electron microscope (EM) are still in use today. These involve fixing the sample in some way by crosslinking the proteins, for example with glutaraldehyde, binding heavy metals (uranium, osmium, and lead) differentially throughout to provide electron contrast, embedding the sample in some kind of plastic, and generating ultrathin sections using a microtome equipped with a diamond knife. These sections, <100 nm thick, are laid out on an electron transparent support of some kind in order to be introduced into the transmission electron microscope (TEM). For the first C. elegans connectomics work, these methods were optimized to promote visualization of cell membranes (so cell processes could be traced through a serial stack) and synaptic structures (presynaptic densities and gap junctions), while not heavily staining cell contents such as ribosomes that obscured these structures. Briefly, worms were cut in half and fixed in osmium tetroxide before embedding, sectioning, and poststaining with uranyl acetate and lead citrate (White et al. 1986).

The electron microscopist who developed this methodology for C. elegans connectomics in the late 1960’s and early 1970’s, Nicol Thomson, was a Scotsman described by Brenner as “a man of great skill!” (Brenner 2005). His results so far surpassed anything subsequent workers could do that 30 years later his images were still valuable and were used for the next round of C. elegans connectomics, which covered the male tail and provided a new digitized reconstruction of the hermaphrodite (Jarrell et al. 2012; Cook et al. 2019). For a reconstruction of the male head, not previously covered, the only advance made in EM was use of a digital, automated TEM (Cook et al. 2019). Perhaps the greatest challenge for connectomics by these methods is handling the delicate sections without loss. Any loss is a potential disaster because neurites might then not be traceable through the stack. The longest series even Thomson could make covered only half a worm (∼10,000 sections at 50–100 nm thickness, ∼0.5 mm). But further, the images obtained by his methodology are somewhat unsatisfying in that much cell ultrastructure is lost and there are gaps throughout separating neural processes that are probably fixation artifacts.

Beginning in the first decade of the 21st century, several laboratories wishing to determine connectivity in nervous systems larger than that of C. elegans, and knowing that this could not be done with the traditional methods, turned their attention to developing new approaches. They addressed all areas: ability to obtain long unbroken series, speed of image acquisition, and speed of image annotation. The new work on C. elegans that took advantage of the existing images focused on speeding up annotation by employing a PC (Xu et al. 2013). Otherwise, the new methods, developed initially for Drosophila and mouse, have only just begun to be exploited for C. elegans (Witvliet et al. 2020). There is still no complete connectome for a single whole animal; the “whole animal” connectomes that have been published, for the male and hermaphrodite, were pieced together from parts of multiple animals and even required conceptually filling in gaps remaining in repetitive regions of the ventral cord where EM images have never been collected (Cook et al. 2019). But the new methods hold great promise and should be employed to the extent possible by anyone contemplating entering this field. For example, consult Scheffer et al. (2020) to see how the latest advances were employed to obtain a Drosophila hemibrain connectome. The following is a guide to some methods recommended for any new work.

Generating images:

Comprehensive descriptions of EM methodology for C. elegans have been provided by Hall et al. (2012) and Mulcahy et al. (2018) and are also available at WormAtlas.org. What follows is a summary of recommended best practices and references for the latest methods and developments. It is advisable for anyone embarking on a project to consult the active laboratories for their latest recommendations and protocols.

Sample preparation:

High-pressure freezing and freeze substitution (Hall et al. 2012; Mulcahy et al. 2018) (WormAtlas.org). This method preserves cell ultrastructure better than the earlier methods, including clear chemical synaptic and gap junction structures (Figure 4A).

Figure 4.

EM imaging and annotation. (A) An image generated using the latest methodology and a scanning EM at 1 nm/pixel resolution (Mulcahy et al. 2018). The sample was prepared by high-pressure freezing and freeze substitution with glutaraldehyde, tannic acid, and OsO₄ fixation and staining. Notice the general lack of large spaces between cell membranes. Upper segment: a gap junction, recognized by the apparently stiffly curving region of closely apposed, darkly staining membranes. Lower segment, a chemical synapse defined by a presynaptic density and synaptic vesicles, both dense core and clear. This seems likely to be a triadic synapse, provided the two lateral processes have receptors for the neurotransmitter released into the intercellular space. (B) Skeleton reconstruction with Elegance. Three adjacent sections in the male preanal ganglion are shown. Locations of cell profiles (blue) and synapses (red) are annotated with database object numbers. Two “contins,” chains of continuously running objects, are marked: contin 908 is male-specific sensory neuron PCBL and contin 641 is sex-shared interneuron AVG (notable with clear cytoplasm). Note the convoluted shapes of some processes. At this location, AVG completely surrounds PCBL, which extends a finger inside AVG that runs for 28 sections. (C) Volumetric tracing in TrakEM2 of the same image as the one shown in the center in (B) Green: AVG; orange: sex-shared motor neurons; pink: the remaining sex-shared neurons; blue: male-specific neurons.

Sectioning:

ATUM (automatic tape-collecting ultramicrotome) (Schalek et al. 2011). This machine, developed by the Lichtman laboratory at Harvard, uses a traditional diamond knife ultramicrotome, but collects sections, running automatically, onto a continuously moving tape. The tape is electron opaque, necessitating the use of the scanning electron microscope (SEM) for image acquisition. Unbroken series of an entire worm should be possible with this machine.

EM:SEM

Modern machines run automatically (Hayworth et al. 2015). The latest advance is development of multibeam machines, which should make acquisition of a stack of C. elegans images extremely rapid (Eberle et al. 2015). The newest TEM machines also can operate automatically, and custom machines 10-times more rapid have been developed that have allowed imaging of the entire Drosophila brain (from a series of 7000 35–40 nm sections) (Bock et al. 2011; Zheng et al. 2018).

FIBSEM (focused ion beam SEM), SBFSEM (serial block face SEM):

These are two methods that combine sectioning and image acquisition into a single operation. Their advantage is they reliably and automatically produce unbroken series of very thin sections and generate stacks of images that are perfectly in register, ready for analysis. Sectioning takes place inside the SEM, which images the face of the sample block after each removal of the previous layer, either by a focused ion beam (FIBSEM, Knott et al. 2008) or with a diamond knife (SBFSEM, Denk and Horstmann 2004). The most recent, massive connectomics dataset, that of the Drosophila hemibrain, was obtained by the FIBSEM method, chosen because the very thin sections (<10 nm) that allow an isotropic (<10 nm in x, y, and z) image volume and perfect alignment facilitate tracing the finest processes (<15 nm) and significantly improve the accuracy of the initial computer-generated segmentation (Scheffer et al. 2020). (The x, y resolution is lower than is possible with SEM as a compromise for speed; to obtain images of higher resolution the SEM is too slow for connectomics.) One limitation of these methods is they are destructive—each section can be imaged only once. Second, with FIBSEM, the possible sample size is small, smaller than a worm. To overcome this limitation for the fly brain, a method was developed for cutting the sample into suitably sized pieces that can be separately imaged and the images accurately montaged back together (Hayworth et al. 2015). Furthermore, they do not allow for the scoring of gap junctions (Scheffer et al. 2020). So far they have not been exploited for C. elegans connectomics but should be considered.

Annotating EMs:

As difficult as generating serial EM images of long intact series is, extracting the connectivity data from the images is the most time-consuming step of a connectomics project, by a factor of 10. In fact, it can never be completed down to every single finest process and synapse, even in C. elegans. In the Drosophila hemibrain reconstruction, somewhere around 50% of the synapses could be scored. Perhaps surprisingly, nonetheless the connectome is nearly complete because the large number of synapses between each pair of connected cells ensures a connection is unlikely to have been completely missed (Scheffer et al. 2020).

Segmentation requires tracing the neural processes from section to section, somehow determining the identity of each process, and annotating the locations of synapses and the identities of synaptic partners. While heroic efforts have been made by computational scientists to teach computers to perform this task, the very high level of accuracy required in order to get a meaningful reconstruction, along with the vast level of often obscured details in EMs, has defeated their stand-alone application. However, they do provide a useful first draft. For the Drosophila hemibrain, automated segmentation of both processes and chemical synapses was followed by 50 person-years, of manual error correction. Apparently, images of C. elegans material are particularly difficult in this regard (Mulcahy et al. 2018; S. Seung, personal communication). Fortunately, the worm is small enough to allow projects to be completed by fully manual annotation. For example, reannotation for skeleton diagrams of previously annotated micrographs covering the hermaphrodite nerve ring required only a few person-months (Cook et al. 2019). There are now several apps for manual scoring of EMs for connectomics. Helmstaedter and Mitra (2012) have provided a review of some earlier packages. Below are programs that have been more recently employed.

Elegance ( https://github.com/Emmonslab/Elegance ) (Figure 4B). Elegance was initially developed to make it possible for digitized legacy C. elegans EMs to be manually annotated on the computer screen with a PC (Xu et al. 2013). It was used for the reconstructions presented in Jarrell et al. (2012) and Cook et al. (2019, 2020). Designed for speed and simplicity, Elegance allows tracing neurites as skeleton tracks of single points (mouse clicks), scoring of synapses and synaptic partners, and generates the resulting maps and connectivity matrices. Images of several adjacent sections, typically three, are displayed simultaneously side-by-side on the screen, a configuration that may be advantageous for identifying corresponding profiles in adjacent sections (S. Cook, personal communication).

As C. elegans is a worm and most neural processes run close to orthogonal to the typically transverse plane of section, making en passant synapses with their neighbors, it is possible to obtain a reasonable estimate of the size of each synaptic object (presynaptic density or gap junction) by simply counting the number of serial sections they run through. By summing over the usually multiple synapses between cells, a weighted adjacency (connectivity) matrix providing the total physical strength of each connection is obtained (see Figure 5B). In current C. elegans data, with section thickness in the 50–100 nm range, synapse size measured this way ranges from a single section to ∼100 sections. The information may be useful for some types of analyses. It may be helpful to eliminate the great many one- and two-section connections, which are more variable and may obscure or complicate a result. The stronger connections are created not only by more synapses but also by larger synapses. If synapse size is not taken into account, the physical strengths, and by proxy the activity strengths, of the strong connections will be underestimated [see extended data Figure 2d in Cook et al. (2019)].

Figure 5.

Levels of structural description A. The adjacency matrix of chemical synapses for an entire animal nervous system, that of the adult C. elegans male (Cook et al. 2019). With a rational ordering of rows and columns it can be seen that the pharynx contains an almost wholly isolated nervous system, that there is a general flow of connectivity towards the end organs, that innervation of the bodywall muscles and the motor neurons that drive them are organized in chains, and that the male has a separate, highly cross connected region, which contains the circuits for copulation. B. A portion of the adjacency matrix showing the values for connectivity of some amphid sensory neurons onto interneurons (in this case, number of EM serial sections); C. A portion of the force-directed layout for the hermaphrodite nervous system covering some of the same neurons as is shown in B (black arrows, chemical connections; red lines, gap junctions). Notice that left/right homologs are placed close together by the algorithm, reflecting the similarity of their connectivity. D. A portion of the synapse table for the hermaphrodite ((Cook et al. 2019)), Supplementary Information 3; also available at WormWiring.org). The fraction of polyads, 16/20 (80%) is typical for the nervous system as a whole. E. Skeleton diagrams of a subset of neurons in the nerve ring and ventral ganglion of the adult hermaphrodite, from WormWiring.org. Balls of three colors indicate chemical input, output, and gap junction synapses, the size reflecting the size of the synapse. These maps are linked to the electron micrographs so any synapse can be inspected by clicking on the synapse on the map. F. Volumetric reconstruction of an L4 hermaphrodite nerve ring. Neurites are assigned colors based on spatial domains obtained from mathematical clustering of the L4 and adult contactome graph (Brittin et al., 2021). These spatial domains are robust to individual variability and support local circuits for modular information processing. (Image: C. Brittin).

Image registration:

Separated EM sections generated with the diamond knife for imaging by either TEM or SEM are physically deformed during the process. The stacks of images through a series therefore cannot be well-aligned across entire sections if only zoom, rotation, translation, and shear are employed—affine or linear transformation. Without further nonaffine (elastic and nonlinear) registration, that is, registration that allows for adjusting pixel locations differentially across an image, skeleton tracks or volumetric maps of neurites will lack a natural smoothness. Software is available for accomplishing a nonaffine alignment (Saalfeld et al. 2012; FijiBento alignment package https://github.com/Rhoana/FijiBento) (Joesch et al. 2016). These must be applied separately before reconstruction with Elegance, but are incorporated as part of other reconstruction packages (see below).

TrakEM2, FIJI:

TrakEM2, (http://repo.or.cz/w/TrakEM2.git) (Cardona et al. 2012), designed initially for Drosophila data, is a far more feature-rich application than Elegance. While it allows tracing neurites as skeletons, it also provides tools for painting profile areas from which volumetric reconstructions can be made (Figure 4C). It allows handling very large datasets by tiling images and swapping tiles in and out of memory as required. TrakEM2 is run as a plug-in for ImageJ and thus has access to the vast array of image-processing tools available in this long-standing and highly developed image processing program. A version of ImageJ with TrakEM2 installed is available as a package called FIJI (Fiji Is Just ImageJ) (Schindelin et al. 2012). TrakEM2/FIJI was used for the volumetric reconstruction of the nerve ring in Brittin et al. (2021). TrakEM2 in FIJI can be employed collaboratively over the web using the app CATMAID (Collaborative Annotation Toolkit for Massive Amounts of Image Data) (Saalfeld et al. 2009; Schneider-Mizell et al. 2016). Current versions of CATMAID allow only for skeleton reconstruction. It was used for skeleton reconstructions by (Witvliet et al. 2020).

Volume Annotation and Segmentation Tool:

Volume Annotation and Segmentation Tool (VAST) is a tool for manual segmentation for volumetric reconstruction that has been developed and used for reconstructing a volume of mouse neocortex (Kasthuri et al. 2015; Joesch et al. 2016). It allows for collaborative annotation of images online. VAST was used by Witvliet et al. (2020) for volumetric reconstruction. It was combined with an automated first draft that took advantage of a prior skeletonization to identify cell boundaries, followed by manual correction with VAST (Mulcahy et al. 2018).

webKnossos:

webKnossos is a tool designed to allow multiple users to trace skeletons as rapidly as possible through mammalian tissue (Boergens et al. 2017). It is designed for images generated by SBFSEM. The SBFSEM method yields isometric voxels—that is, resolution in the x, y plain of each image (15 nm as used here) is the same as in the Z direction, section thickness. Such images, already perfectly aligned, make it possible to reconstruct a 3D image space first and then carry out the reconstruction in that rotatable volume. Annotators can employ a “flight mode” in which they follow along and mark a process as it progresses through the volume.

The reliability of connectomics data:

In spite of some two decades of developments since EM-based connectomics was taken up with modern computer assistance, determining connectomes remains an inexact enterprise. Significant uncertainties affect all three aspects: process tracing, chemical synapse annotation, and gap junction annotation.

Beyond problems arising from thin processes running in the plane of section and poor image quality, process tracing is confounded by the complex shapes of neural processes (Figure 4, B and C). As in other animals, C. elegans processes often extend tiny protrusions, branches, or spines, dubbed “twigs” in the Drosophila literature [see Cook et al. (2019) extended data Figure 1A]. These can make synapses. Depending on section thickness and neurite orientation, such branches may appear in the images for just a few sections before disappearing again and it may be impossible to determine which of the adjacent processes they emerged from.

For chemical synapses, the difficulty on the presynaptic side is that the presynaptic density can have a range of sizes, from perfectly robust and obvious down to a tiny smudge of great uncertainty in a single section. The clarity is strongly dependent on the fixation and staining conditions.

On the postsynaptic side, the difficulty is deciding which cells are the true postsynaptic partners, since in C. elegans there is, in general, no visible postsynaptic structure. The only criterion to use is proximity to the neurotransmitter release site defined by the presynaptic density. Different scorers will make different judgements in this regard (Xu et al. 2013). One rule to follow might be that the postsynaptic membrane should overlap the site of the presynaptic density. In many examples, the presynaptic density clearly faces onto multiple cells. But in other examples, the synapse shown in Figure 4A is one, nearby cells facing onto the intercellular space into which the neurotransmitter is released would be postsynaptic if they expressed the appropriate receptor, even though they may not actually overlap the presynaptic density. In many cases, it appears as though a neuron extends a “finger” towards a synapse as if wishing to be postsynaptic. Multiple such fingers may crowd around a release site without actually “touching” the presynaptic density. There is no way to resolve the uncertainty from the EM structure alone.

Gap junctions are even more difficult to score and their visibility even more dependent on the conditions for preparing the sample. While some are unambiguous, such as the one shown in Figure 4A, there is a continuous spectrum down to structures of complete uncertainty that are rejected by most, but not all, annotators. Due to concern over the large number of uncertain structures being scored as gap junctions during the reconstruction of the male posterior nervous system, their distribution was analyzed (Jarrell et al. 2012; Emmons laboratory unpublished). They were found to be nonuniformly distributed along processes, some regions of a given process having many, other regions none. In bare regions, the process contacted just as many neighbors as in regions with many. The evidence favored the interpretation that these are biological structures, not EM artifacts. But whether they are gap junctions rather than some other type of adhesion is unknown. A recommendation to annotators is to score everything, since the process of examining the images is so time-consuming one does not want to do it twice. But record a degree of certainty with each object. Later, structures scored as uncertain can be eliminated from an analysis if desired.

Websites for EM data:

WormWiring.org (https://wormwiring.org/) hosts C. elegans connectomics data, ranging from whole-animal adjacency matrices to skeleton and volumetric maps of single neurons and lists of individual synapses, their sizes and partners. The data include a unique identifier for each scored synapse; clicking synapses on maps brings up the EM where it was scored. Network diagrams are available in Cytoscape that allow exploration of connectivity of individual neurons. An interactive PDF version of Cook et al., Figure 1 (Cook et al. 2019), Supplementary Information 1 A3 https://doi.org/10.1038/s41586-019-1352-7, is available that allows correlation of nodes in the network with cell body anatomical locations and brings up additional information such as neurotransmitter and links to information at additional websites (WormBase.org; WormAtlas.org).

Nemanode.org is a tool designed to explore connectivity of individual neurons, including neurotransmitter usage. It draws data from the multiple reconstructions of Witvliet et al. (2020).

A digitized collection of a large number of EMs, including most of the mostly unpublished micrographs from the Brenner group, is available at WormImage.org (https://wormimage.org/).

Identifying synapses and determining connectivity by fluorescence labeling

EM is technically limited to providing a static shot from only a very few, in many projects only a single, individual animal. This raises the question of the generality of the results obtained and moreover makes it essentially impossible to carry out any experiment on synapse formation. To overcome these problems, fluorescent probes that reveal both chemical and gap junction synapses have been developed for cell-specific expression from transgenes or, more recently, from the endogenous gene locus itself using CRISPR/Cas9 editing.

Chemical synapses:

To visualize chemical synapses, a variety of fluorescently tagged proteins that localize to pre- or postsynaptic specializations or structures have been employed. These have included proteins that localize to synaptic vesicles (SV) or the presynaptic active zone and postsynaptic neurotransmitter receptors. Synapse localization was first carried out via immunostaining of presynaptic proteins such as the vesicular calcium sensor synaptotagmin (SNT-1; Nonet et al. 1993), and later using GFP-fusion transgenes of proteins such as the vesicular SNARE complex protein synaptobrevin/VAMP (SNB-1; Jorgensen et al. 1995; Nonet 1999). The fluorescently tagged SV-associated GTPase RAB-3 (Mahoney et al. 2006) is also commonly used, although as a nontransmembrane protein it can occasionally become dissociated from SVs. In general, due to their abundance, fluorescently tagged SV proteins generate a bright signal that can be easily visualized on a variety of microscopes. However, because SVs can fill presynaptic compartments (and be continuously transported between them), resolving individual synapses using SV markers can sometimes be a challenge.

The first active zone protein utilized was the protein SYD-2/Liprin-α identified in studies that demonstrated its localization to a sub-region of the presynaptic compartment surrounded by fluorescently tagged SVs (Zhen and Jin 1999). Development of a GFP-tagged SYD-2 marker (Yeh et al. 2005), Identification of genes involved in synaptogenesis using a fluorescent active zone marker in Caenorhabditis elegans allowed for its use in genetic screens to identify mutants with aberrant synapse development. Additional active zone markers (e.g., SYD-1, ELKS-1, UNC-10/RIM, UNC-13, and so on.) have since been developed and used for characterizing various mutants. More recently, a novel active zone protein with homology to vertebrate Piccolo, Clarinet (CLA-1), was shown to localize very precisely to active zones (Xuan et al. 2017). The more discrete localization pattern of these active zone proteins makes them more amenable for discrimination and quantification of individual synapses than proteins of SVs. In addition, the advent of super-resolution imaging techniques, including single-molecule localization microscopy, has led to the interrogation of multiple presynaptic proteins at the level of individual active zones (Kurshan et al. 2018).

There is no postsynaptic density in C. elegans and there is no broad method for labeling the postsynapse comparable to the methods available for the presynapse. In early attempts to visualize postsynaptic neurotransmitter receptors, the levamisole-sensitive acetylcholine receptor subunit UNC-29 and the GABA receptor subunit UNC-49 were tagged with GFP and shown to cluster at putative postsynaptic sites in both neurons and muscles (Fleming et al. 1997; Bamber et al. 1999). Visualizing clusters of the glutamate receptor GLR-1 within neuronal processes required targeting the fluorophore to a region of the protein not required for its localization, and was first achieved by insertion of GFP within the C-terminal tail, just before the last 16 amino acids (Rongo et al. 1998). Likewise, visualizing postsynaptic clusters of the nicotinic acetylcholine receptor subunit ACR-16 required targeting the GFP to an intracellular loop (Francis et al. 2005). Advances in light microscopy techniques have led more recently to the identification of postsynaptic acetylcholine receptor clusters localized to the tips of dendritic “spines” in a subset of neurons (Philbrook et al. 2018; Cuentas-Condori et al. 2019).

The advent of CRISPR/Cas9 transgenesis has led to an explosion in the development of synaptic proteins fluorescently tagged at the endogenous gene locus. Cell-specific labeling techniques (e.g., Schwartz and Jorgensen 2016; He et al. 2019) allow for the visualization of such proteins in individual neurons. Overall, the subcellular localization of endogenously tagged synaptic proteins has largely mirrored that of the previously used, typically over-expressed products of transgenes. However, subtle differences in parameters such as cluster size do exist for several proteins, particularly those with many protein-protein interaction domains that make them susceptible to aggregation when over-expressed.

Going beyond separately labeling pre- or postsynaptic structures, to specifically label a synaptic connection between two neurons, two methods have been developed. In GRASP, fragments of GFP are expressed respectively in pre- and postsynaptic neurons and yield a fluorescent signal only if they are brought together at a synaptic connection (Feinberg et al. 2008). In iBLINC, bacterial biotin ligase birA is displayed on the presynaptic membrane and biotinylates an acceptor peptide displayed on the postsynaptic side if these are brought together at a synapse (Desbois et al. 2015). The biotinylated protein is then detected by binding to fluorescently labeled streptavidin secreted from coelomocytes into the extracellular compartment. The well-studied cell adhesion molecules neurexin and neuroligin are used to localize the enzyme and acceptor peptide to pre- and postsynaptic cells, respectively.

Published results on connectivity obtained by these single-synapse methods have been encouragingly consistent with EM when taken as an average across multiple individuals. However, the number of fluorescent puncta, indicating synaptic specializations or synaptic connections, that are observed in individual animals varies widely, with many individuals seemingly having connectivity quite inconsistent with EM scoring. It’s possible that these are true individual differences, but this seems unlikely as the outliers are sufficiently frequent to expect that at least one of the animals reconstructed by EM would have been such an outlier, considering the relatively large number of connections (>20) that have been checked by fluorescence. So far, there have been no inconsistencies, suggesting this could be a problem with the fluorescence data due to transgene variability. These methods are relatively new and a number of issues remain unanswered. Will neurexin and neuroligin localize to any synapse? Can these synapse-labeling constructs cause apparent synaptic connections? Can they stabilize a connection that otherwise is labile? Nevertheless, they have proven valuable and reliable so far, for example documenting sex-specific pruning of synapses during development (Oren-Suissa et al. 2016).

Gap junctions:

Gap junction connections have been visualized by attaching fluorescent labels to innexin gap junction proteins (Starich et al. 2009; Meng et al. 2016; Bhattacharya et al. 2019). Proteins that have been used include UNC-7, UNC-9, INX-6, and CHE-7, with GFP, RFP, or YFP attached at N- or C-terminals. Apparently, such tags do not interfere with either gap junction assembly or function.

Visualizing and analyzing the C. elegans connectome and the structures of neurons and synapses; graph theory

The papers on the C. elegans connectome at the time of this writing are (White et al. 1986; Hall and Russell 1991; Jarrell et al. 2012; Cook et al. 2019, 2020; Witvliet et al. 2020). Volumetric reconstruction of the hermaphrodite nerve ring is provided by Brittin et al. (2021) and Witvliet et al. (2020). Connectivity in the nematode species Pristionchus pacificus is given by Bumbarger et al. (2013).

These papers describe the structure of the nervous system across a number of levels, all of which are relevant for understanding its development and function (Figure 5). A connectome can be described by a connectivity or adjacency matrix that gives the total amount of connectivity, chemical and gap junctional, between pairs of cells. This is a highly abstracted description amenable to mathematical network analysis, where cells are reduced to nodes (vertices) in a network and connections, often due to multiple synapses, are reduced to single edges (weighted if something about the strength of the connection is known)—it includes no information beyond this about the physical structure. The physical structure is described by the locations, volumes, and neighbors of individual neurons, the locations where they form their synaptic connections, and the structures of the individual chemical and gap junction synapses, including the combinations and arrangement of the multiple postsynaptic partners at polyadic chemical synapses. Ultimately, modelers will need to combine information across all these levels, starting from the electrical conductance of the neurons and their membranes and the properties of their individual synapses and maps to the overall architecture of the network in order to provide a complete description—one from which behavioral output may be predicted from the overall structure and the properties of individual cells.

Adjacency or connectivity matrices and network diagrams:

The adjacency matrix, is simply a table where nodes and columns list the neurons and end organs, and the values in the table show which ones are connected (Figure 5, A and B). The term is from graph theory (see below), where connected nodes in a network are said to be adjacent in the network. Typically, for chemical connections, the rows are the presynaptic neurons and the columns are the postsynaptic neurons and end organs. The matrix of gap junction connections is symmetrical assuming the gap junctions are nonrectifying. The values in the table may be binary (Boolean), indicating the presence or absence of a connection, or weighted, giving the strength of the connection (either number of synapses, or, if synapse size is taken into account, a measure of the total morphological strength of the connection) (Figure 5B).

If the rows and columns are ordered in the right way, a view of the overall organization of the network is revealed. For the C. elegans connectome, if the first entries are the sensory neurons, then the interneurons that receive sensory input, then the targets of those interneurons, and so forth, ending with the end organs, the progressive flow of information through the network can be seen (Figure 5A).

An even more visual representation can be obtained by using a network layout program that draws a 2D diagram (Figure 5C). By means of an algorithm that clusters or more widely separate nodes according to whether they are strongly, weakly, or not at all connected, a map of the overall connectivity is obtained. Such a map for the C. elegans nervous system arranges the nodes in clusters that reflect the neuroanatomy of the worm ganglia, demonstrating the developmental importance of minimal wiring length (Cook et al. 2019). Pathways of information flow can be visualized from sensory neurons or sex-specific parts of the nervous system through interneuron pathways to the motor system.

Available layout programs include Cytoscape (cytoscape.org) and Gephi (gephi.org). Both are well-developed programs that provide a wide array of visualization and graph analysis tools. A particularly useful type of layout is a so-called “spring electric” or “force-directed” layout, where nodes are modeled as being simultaneously repelled from each other, as if they carried a like electric charge, and attracted to each other by a spring whose strength is proportional to the strength (weight) of the connection between them. This network is computationally allowed to “relax” to a minimal energy configuration. For a complex network like a nervous system, a true minimum cannot be found within a reasonable amount of computational time and the calculation is stopped after some threshold is reached. The result is a somewhat different layout each time the program is run but it is always consistent and revealing. The spring electric algorithm of Allegro was used in Cytoscape to make Figure 1 of Cook et al. (2019).

Another way to generate an informative 2D layout is to arrange the nodes according to their hierarchical relationships. This was first done for the C. elegans chemical connectome by Richard Durbin, who wrote a program that arranges the neurons from top to bottom such that the maximum number of chemical synapse arrows point downwards (Durbin 1987). Such an analysis gives a satisfying layout with sensory neurons at the top and motor neurons at the bottom. But it also shows a few interneurons surprisingly high in the layout. These have significant “feedback” output onto sensory neurons. Since that time, many algorithms for studying hierarchy or directionality in networks of all kinds have been published. The one used for C. elegans by Varshney et al. (2011) and Cook et al. (2019) is that of Carmel et al. (2004).

Structure of chemical synapses:

Neither adjacency matrices nor 2 D layouts show the polyadic structure of the chemical synapses. This information should be carefully recorded during annotation and can be provided as lists of individual synapses and their properties (Figure 5D) or on neuron maps. The significance of multiple postsynaptic cells has yet to be analyzed in the context of either development or function. Possibly induction of synapse formation by signals from more than one postsynaptic cell plays a role in a combinatorial code for connectivity. Alternatively, polyads might have been favored for reasons of economy or synchrony of function.

Skeleton maps:

Neuron skeleton maps describe how neurons run through the animal and where synapses are located (Figure 5E). At this most basic level of information about the physical structure, the locations of the synapses along the length of a process are obtained—showing whether there are distinct input dendritic and output axonic regions, whether there is clustering or interspersion of connections to particular other cells, and the identities of the synaptic partners. The maps in the first C. elegans connectome publication were mostly skeleton diagrams (White et al. 1986).

Volumetric reconstruction: neighborhoods:

Whereas skeleton maps are generated from tracks of points placed through a stack of EMs (Figure 4B), volumetric reconstructions are generated by “painting” the neuron profiles, a much more laborious process (Figure 4C). Such a fuller reconstruction provides neuron caliber, an important feature for modeling conductance. It also makes it possible to determine the neighbors of each neuron, that is, the other neurons that its membrane appears to contact. Two programs are available for extracting a matrix of cell contacts from tracings made in TrakEM2, written respectively by Brittin et al. 2021) and A. Bloniarz (WormWiring.org).

As a neuron can only synapse onto another neuron that it makes contact with, how the neighbors of a neuron are specified during nervous system development is important to understand in order to understand the process of synapse formation. The C. elegans nervous system is noteworthy for consisting in part of mostly unbranched neurons. White noted that the limited number of neighbors shared by unbranched neurons had an important implication for synapse formation by severely limiting the number of possible synaptic partners (White et al. 1983; White 1985). While this is the case for the nerve ring and the various nerve cords of both sexes, it is not true for the major neuropil of the male posterior nervous system, the preanal ganglion, where the neurons have more typical, branchy structures (Emmons 2016). Finding the correct neighborhood is essential, but it is not sufficient as many neurites travel together for long distances without forming synapses. Closer analyses have confirmed the lack of correlation of synapse formation with amount of contact (Durbin 1987; Brittin et al. 2021). Nevertheless, reconstructions of the nerve ring have revealed that the structure is highly organized with respect to functional pathways (Brittin et al. 2021; Moyle et al. 2021). Neighborhood selection is clearly one necessary step in the process of establishing the connectome.

Graph theory—analyzing network architecture:

Connectomic reconstructions now in a number of animals have revealed that nervous systems contain networks of connections. While very large, complex networks of related entities have always been known (think of the relationships between organisms in an ecological community, as an example), new ones keep arising (think of the internet) and the increasing capabilities of computers to describe and analyze them has made understanding their emergent and common features a problem of great theoretical and practical interest. Since its first description in 1986, the C. elegans connectome has made an important contribution to these studies by providing theorists with an example of a natural neural network. Its analysis has contributed to development of such fundamental concepts as small world, motifs, and modularity (Watts and Strogatz 1998; Milo et al. 2002; Newman 2006; Figure 6). The topological features of nervous system networks that have been revealed by connectomics reconstructions, including that of C. elegans, have been compared to features of artificial neural networks that lead to computational structures such as the perceptron (Rosenblatt 1958) and attractor, Hopfield networks (Hopfield 1982; Jarrell et al. 2012). How far this analogy will hold, whether brains and very large, computer-based AI neural networks that can play GO really work the same, remains to be demonstrated.

Figure 6.

Some important properties of natural networks. (A) Motifs. The frequency of 1-, 2-, and 3-node motifs in the C. elegans connectome (blue dots) compared to the frequency in a set of computationally created random graphs generated with the same degree distribution by randomly exchanging the endpoints of the edges (red plusses, +) (from Cook et al. 2019). The so-called “feedforward loop” (motif 10) was found to be present at significantly high frequency in the C. elegans connectome (Milo et al. 2002). Note that here, in motifs 10, 12, 14, and 16, two nodes connected to a third are also connected to each other, accounting for the high clustering coefficient of the connectome graph. Prevalence of such small motifs suggest local computational functions. (B) Modularity, hubs, rich clubs. A simple diagram illustrating modules, hubs, and rich clubs from Bullmore and Sporns (2012). Hub neurons lie on many shortest paths between neurons. If hub neurons are connected to each other (right), they constitute a rich club. Discovering mathematically modules or communities in networks has been a major field of enquiry. In analysis of the C. elegans male mating circuits, discovery of modularity elucidated function, allowing groups of neurons and muscles to be associated with the several steps of the mating program (Jarrell et al. 2012). Several of the major interneurons in the sex-shared nervous system that control the bodywall motor system are hubs and, being also connected to each other, constitute a rich club (Towlson et al. 2013). A force-directed layout algorithm places these neurons in the center of the network diagram (Figure 1, Cook et al. 2019) (C). Small World. Regular graphs (left) have a high clustering coefficient (note that in this example C = 1) but long average minimum path length. Random graphs (right), where all nodes are connected with the same probability, have short average minimum path lengths but low clustering coefficients. It was discovered that by adding just a few long-range connections to a regular graph (middle), the average minimum path length fell precipitously while the clustering coefficient remained high, a property dubbed “small world” (diagram from Watts and Strogatz 1998).

The field of mathematics that focuses on the properties of networks is known as graph theory. The subject has been developed primarily since the mid-20th century. Mathematical graph theory comprises the study of structures consisting of relationships between pairs of objects. A “graph” as typically plotted by scientists and engineers to display relationships between measurements or a mathematical functional relationship may be considered a special case of a mathematical graph. In this typical scientific graph, each X is associated with a single Y. In the generalized form, each “X” may be associated with multiple “Y”s, and vice versa, in other words, it describes a network. In graph theory, which uses the language of set theory, the set of paired relationships is called a graph, the paired elements are known as vertices or nodes, and the relationships between them are known as edges or links. Papers using graph theory uniformly start with the statement: G = (V, E), where V = (V₁, V₂, _…) and E = (V₁V₂, V₁V₃, _…)—“the graph G consists of a set of vertices, V, and a set of edges, E.” The graph is a directed graph if the pairs of vertices are ordered and it is unweighted or weighted depending on whether the edges are Boolean or have scalar values.

The properties of networks derived from graph theory have been described in many places (see Box for definition of some of the terms that may be encountered in the C. elegans literature). The graph of the C. elegans connectome is a sparse graph, meaning most (90%) of the vertices are not directly connected to each other, but it nevertheless comprises a single component, meaning an unbroken path can be found connecting, directly or indirectly, any pair of vertices.

Box: Some Graph Theoretic and Network Theory Terms that May Be Encountered in the C. elegans Connectomics Literature

Betweenness centrality: extent to which a vertex lies on the shortest paths connecting other vertices

Clustering coefficient: the probability that if two vertices are connected to a third, they are connected to each other

Connected component: a set of vertices within a graph where each vertex has a path to all the others

Degree: property of each vertex: the number of edges attached to it

Degree distribution: the distribution of vertex degree values over a graph

Diameter: the length of the longest minimum path between two nodes

Directed (undirected) edge: the pairs of nodes are ordered (unordered)

Directed (undirected) graph: a graph made up of directed (undirected) edges

Edge, link, connection: equivalent terms

Graph: a  mathematicalstructure consisting of a set of objects (vertices) and pairwise relations between them (edges)

Hub: a node having a far higher number of edges or links (degree) than the average for the graph

In degree, out degree: the number of directed edges that point in to or out from a node

Module, community: (equivalent terms) a set of vertices more strongly connected to each other than to vertices outside the community; more strongly connected to each other than expected for a random graph with the same degree distribution.

Motif: a  patternof connections between a small number of nodes (2-5) that recurs with statistically high frequency when compared to the frequency in a random graph with similar properties (e.g.,degree distribution)

Path length: t h e number of edges in a path

Path: a set of  edgesconnecting two nodes through other nodes

Random graph: A graph generated by randomly connecting all pairs of vertices with equal probability

Rich club: a set of hub (high-degree) nodes connected together by short paths

Small world: a property of many natural sparse graphs such that they have a high average clustering coefficient and also short average minimal path length between  pairsof vertices

Sparse graph: a graph in which most pairs of vertices do not have an edge connecting them

Vertex, node: equivalent terms

Several network properties that are applicable to the C. elegans connectome are shown in Figure 6. While most graph theoretic concepts are intuitive, it was considered surprising when it was found that most natural networks have a property known as small world (Watts and Strogatz 1998). Small-world networks are sparsely connected, yet even if they are very large, any pair of nodes can be connected over a small number of links, a short path—six degrees of separation. Three networks used to demonstrate this property were the C. elegans connectome network, the power grid of the western United States, and the collaboration graph of film actors. Network properties of the C. elegans connectome that have been analyzed include its modularity, hierarchical modularity, and consequent efficiency of information processing and wiring cost (Chen et al. 2006; Bassett et al. 2010), existence of hubs and rich club neurons (Towlson et al. 2013), and controllability (Yan et al. 2017). A good place to start reading about graph theory is Wikipedia. Papers that have summarized graph analysis of connectomes include Bullmore and Sporns (2009) and Van den Heuvel (2016). A number of software packages are available for analyzing the properties of graphs, including functions within Cytoscape (Smoot et al. 2011), igraph (Csardi and Nepusz 2006), and for use with MATLAB (Rubinov and Sporns 2010; Konganti et al. 2013). A. Bloniarz has provided some simple MATLAB routines for analyzing C. elegans adjacency matrices, available at WormWiring.org.

Making comparisons:

An important issue arising in the analysis of connectomes is that of the similarity of the nodes. For example, it was helpful to determine for C. elegans how similar is the connectivity of neurons thought to be of equivalent cell type, for example, left/right homologs of paired sensory neurons. The amount of such similarity and dissimilarity gives an indication of the reproducibility of the developmental program that determines the wiring (Jarrell et al. 2012; Cook et al. 2019). The problem of comparing two graphs to identify their similarities and differences encompasses an entire field of enquiry, known as graph matching (consult Wikipedia). A number of standard approaches to finding similar nodes in a network are unsatisfactory for connectomics data because the formulas treat an edge weight (determined from number of EM sections or number of synapses) of 0 as very different from an edge weight of 1. But connections identified by only a single synapse seen in a single EM section are of doubtful significance, not considered significantly different from 0. Bloniarz devised a satisfactory new approach that takes this consideration into account (included as part of his MATLAB routines available at WormWiring.org) (Jarrell et al. 2012).

Neuronal Ca²⁺-imaging for assessing neuronal activity

Imaging of neuronal activity mostly via measurements of intracellular Ca²⁺ has enabled a large community of researchers to perform neurophysiological experiments in C. elegans. Ca²⁺-imaging vastly expanded the repertoire of sensory and behavioral paradigms under which such measurements can be performed. It enabled investigating information processing from the sub-cellular scale to the level of nervous-system wide network activity as well as directly relating these activities to the instantaneous behaviors of freely moving animals. During the past two decades, Ca²⁺-imaging experiments thereby revolutionized our current understanding of how worms sense their environment, integrate such information with experience and their internal state, and how evoked as well as spontaneous behaviors are generated. This study extends on a previous study in wormbook (Kerr 2006), briefly reviews some of these new insights, focuses on the available techniques, and provides some guidelines to design, perform and interpret Ca²⁺-imaging experiments with C. elegans.

Physiological properties of C. elegans neurons important for interpreting Ca²⁺-imaging experiments

Morphology:

Compared to larger invertebrates and vertebrates, C. elegans neurons have a simple morphology comprised of a soma and mostly just one to two major neuronal processes, many of them have mixed axonal- and dendritic identities, i.e., they contain both pre- and postsynaptic sites. Typically, sensory processes are largely devoid of synapses, like the ones of chemosensory neurons or the putative proprioceptive processes of motor-neurons (White et al. 1986). More diverse morphologies can be found in the endings of sensory neurons (Doroquez et al. 2014), branching patterns of FLP and PVD mechano-sensory neurons that tile the body-surface (Oren-Suissa et al. 2010; Smith et al. 2010; Albeg et al. 2011), synaptic varicosities (White et al. 1986; Cook et al. 2019), and even small spine-like structures in motor neurons (Philbrook et al. 2018; Cuentas-Condori et al. 2019). Neuronal somas are small: ∼2 µm in diameter (compare to the ∼20 µm sized somas of mammalian pyramidal cells); likewise, their processes are short ranging from 10 s of microns to about 1 mm with small diameters of about only 100–200 nm (White et al. 1986; Cook et al. 2019). Most cells are densely packed in ganglia, and tissues are under pressure held by the animal’s cuticle.

Electrical signals:

These geometrical and morphological constraints represent major challenges to perform classical patch clamp recordings from neurons demanding great experimental skills and patience: typically, animals are glued to a substrate and a tiny slit is cut into the cuticle, which lets somas to pop-out making them accessible to recording electrodes. This setup restricts recording-sites to somas, which due to their small size and volume are difficult to stably patch (Goodman et al. 1998; Lindsay et al. 2011).

Despite these technical challenges, invaluable insights into the physiology of C. elegans neurons could be made from such patch clamp recordings. As anticipated from their small dimensions, neurons seem to be nearly isopotential at steady state, meaning that membrane potential (Vm) should be uniform across the entire cell membrane. However, these assumptions were made under additional theoretical considerations (Goodman et al. 1998); since recording from the thin processes is not feasible and considering the complex morphologies described above, functionally relevant local Vm fluctuations cannot be excluded. Consistent with the lack of voltage-gated sodium channels encoded in the genome, early studies did not report classical action potentials. However, C. elegans neurons should not be seen as entirely passive-linear encoders: ASE sensory neurons responded to current injections in a multi-phasic fashion exhibiting an inflection point in the rise of Vm; voltage-gated K⁺ and Ca²⁺ currents were proposed to convey sensitivity and high dynamic range to these neurons (Goodman et al. 1998). In a later study, RMD motorneurons were shown to exhibit Ca²⁺-dependent graded regenerative signals, that were characterized by depolarized plateau-states and bi-stability, and which the authors termed action potentials (Mellem et al. 2008) [see also Lockery et al. (2009) for discussion]. Finally, Liu et al. reported that under certain conditions AWA neurons generate voltage spikes, which conform with the stricter definition of action potentials, i.e., all-or-none, invariant and self-terminating. Voltage-gated Ca²⁺-channels mediate the upstroke phase while potassium channels mediate the downstroke phase of these events (Liu et al. 2018b). Allover, surveying reported I/V-relationships obtained from electrophysiological recordings reveals a diversity of electrical properties: some neuron classes seem to exhibit near linear I/V relationship like AVA, RIM, AVE, PLM (Liu et al. 2018b); some neurons exhibit rectifying properties (AIY) (Liu et al. 2018b) and others bi-stability (ASE, AWA, AFD, AWC, ASH, and RMD) (Goodman et al. 1998; Mellem et al. 2008; Liu et al. 2018b). Lockery and Goodman (2009) discussed that neither graded linear encoding nor action-potential-based rate-coding would be reliable in small neurons like the ones found in C. elegans. This is because they express low numbers of individual ion channels (Goodman et al. 1998) and in small neurons with tiny diameter processes, the stochastic opening and closing of them can have a significant contribution to Vm generating noise (Faisal et al. 2008). It is possible that the diversity of electrical properties found in worm neurons, like active plateau potentials, is a means to overcome these biophysical limitations (Lockery and Goodman 2009).

Synaptic release:

C. elegans has been a long-standing genetic model to study the synaptic release machinery, which is largely conserved between nematodes and larger animals (Richmond 2005). Fusion of SV with the cell membrane, a key step in neurotransmitter release, is Ca²⁺-dependent. Importantly, C. elegans synapses engage in graded release, which has been demonstrated not only at the neuromuscular junction (Liu et al. 2009a), but also at sensory-to-interneuron synapses. Two studies showed graded transfer characteristics from AFD thermosensory neurons to AIY interneurons (Narayan et al. 2011) and from ASH nociceptive neurons to AVA interneurons (Lindsay et al. 2011).

Here, we discuss these cellular and biophysical properties of C. elegans neurons in such detail because they have important implications on interpreting neuronal Ca²⁺-signals. The active and passive currents that underly the graded fluctuations in Vm all receive significant contributions from Ca²⁺-channels. Furthermore, presynaptic Ca²⁺-levels directly link Vm fluctuations to synaptic release, i.e., the output of each neuron. Therefore, intracellular Ca²⁺ serves as a reliable readout of the information processed and transmitted by each neuron. This view, however, should be always taken with some caution. Intracellular Ca²⁺ is not synonymous with Vm, which receives contributions from other cationic and anionic conductances. Intracellular Ca²⁺-homeostasis depends on additional factors like Ca²⁺-pumps, internal stores, and intracellular Ca²⁺-buffers, which operate on longer time-scales than Vm fluctuations. Furthermore, Ca²⁺-currents generally do not directly report inhibition, i.e., a transient decline in a Ca²⁺-signal does not necessarily report inhibitory input to the neuron. Intracellular, Ca²⁺ can act as a second messenger modulating the properties of ion channels, hence Ca²⁺ can fed back on Vm in complex ways. This explains surprising observations in the nociceptive ASH neurons where certain treatments, like serotonin application, had opposite effects on the magnitude of evoked Ca²⁺-signals and their corresponding changes in Vm (Zahratka et al. 2015; Williams et al. 2018).

In summary, although in a largely nonspiking nervous system, Ca²⁺-signals should be interpreted as time-convoluted proxies for graded neuronal activity. In this respect, Liu et al. performed very informative simultaneous current-clamp Ca²⁺-imaging recordings from AWA sensory neurons. While voltage rapidly plateaued in response to current injection, Ca²⁺-imaging signals continuously increased for seconds. Fast Vm fluctuations, including AWA action potentials, were strongly convoluted in the Ca²⁺-signal; however individual action potential could still be computationally decoded from them (Liu et al. 2018b). Below we discuss that other factors like sensor kinetics and Ca²⁺-buffering might have further contributed to this effect. These data, nevertheless, impressively demonstrate how Vm and intracellular Ca²⁺ can relate to each other.

Genetically encoded Ca²⁺-indicators

The well-established technologies for convenient and fast transgenesis in C. elegans together with its transparent tissue are ideal for the combination with genetically encoded indicators of neuronal activity. The introduction of genetically encoded Ca²⁺-indicators (GECIs) triggered an avalanche of studies in worms that was ignited by seminal work in the Schafer lab (Kerr et al. 2000; Hilliard et al. 2002; Suzuki et al. 2003); since then it became feasible, for a broad audience of researchers, to monitor neuronal activity in a cell type specific and noninvasive manner. During recent years GECIs underwent rapid development with iterative improvements that transformed neuroscience in general.

GECIs operate by the principle of linking a Ca²⁺-binding element to FPs, rendering their fluorescent properties sensitive to Ca²⁺-concentrations. GECIs can be classified (1) based on their Ca²⁺-binding mechanism or (2) based on their means to change Ca²⁺-dependent fluorescence.

1. Troponin based indicators, e.g., Tn-XXL, use Ca²⁺ binding domains of muscle Troponin C (TnC) from skeletal and cardiac muscle (Heim and Griesbeck 2004; Mank et al. 2008; Thestrup et al. 2014), in contrast to sensors based on Calmodulin (CaM) combined with the M13 domain of myosin light chain kinase (e.g., GCaMPs) (Nakai et al. 2001; Nagai et al. 2004). The choice between those two types of indicators potentially has implications. Heim et al. discussed that CaM compared to TnC is a more ubiquitous signaling molecule with a plethora of interaction partners, potentially altering cell physiology when over-expressed in neurons (Heim and Griesbeck 2004). Indeed, it was shown that GCaMP functionally interferes with L-type calcium channels (CaV1) in mammalian neurons, an effect that could be mitigated by engineering an apoCaM-binding motif at the N-terminus (GCaMP-X) (Yang et al. 2018). To date, there are no systematic C. elegans studies that compare TnC based indicators or GCaMP-X with other CaM based indicators. However, these alternative probes should be considered, particularly when researchers suspect cytotoxicity in their assays, due to GECI overexpression.

2. Another classification of GECIs is based on the mechanisms by which their fluorescence signals report changes in Ca²⁺. Ratiometric indicators make use of Förster resonance energy transfer (FRET) between a donor fluorophore (e.g., CFP) and an acceptor fluorophore (e.g., YFP), linked via the Ca²⁺-binding domain. The conformational change upon Ca²⁺ binding changes the distance and orientation of the donor-acceptor pair in a way facilitating energy (“virtual photons”) to be transferred between them (Figure 7A). In practice, enhanced FRET upon Ca²⁺-binding typically causes a concomitant decrease and increase in the fluorescence emission of the donor and acceptor respectively. This can be conveniently measured as a ratio signal upon selective donor excitation and using a beam-splitter in the optical pathway. The advantages of FRET-based indicators can be twofold: first, calculating the emission ratio automatically corrects for correlated noise in the two channels, e.g., motion artifacts. However, this comes at a trade-off: uncorrelated noise, conversely, could be vastly amplified by the division step. Second, if properly calibrated, FRET-based Ca²⁺ imaging could yield absolute intracellular Ca²⁺ concentrations. However, to the best of our knowledge, there is no C. elegans study that exploited this opportunity so far. Most nonratiometric GECIs exhibit a circular permutated and thus perturbed FP (Figure 7B). Here, the conformational change upon Ca²⁺ binding reconstitutes the fluorophore leading to a change in absolute fluorescence (Figure 7B). The advantages of single fluorophore indicators are that they comply with a simpler imaging setup using only a single excitation/emission beam-path. Moreover, recent engineering efforts mostly focused on improving the GCaMP family of GECIs, now offering a large set of probes with different properties to choose from Akerboom et al. (2012), Chen et al. (2013), and Dana et al. (2019). Given how the field is leaping, providing recommendations of individual GECI here will probably be outdated soon. However, the following guideline should be considered when generating new transgenes for Ca²⁺-imaging.

Figure 7.

Schematics of GECI working principles. (A) Ratiometric FRET-based indicators. Conformational changes upon Ca²⁺-binding promotes FRET between a donor-acceptor pair (e.g., CFP/YFP variants in yellow chameleons). When the donor gets selectively excited, FRET causes a reduction in its detectable emission and an increase in the emission of the acceptor. (B) Operating principle of GCaMP. Circular permutation of N- and C-terminal halves perturbs fluorescence properties, which get reconstituted upon Ca²⁺-binding. All domains and linkers are not drawn to scale. (C) Hill plots and theoretical brightness levels of three selected GCaMPs. The equivalent brightness of EGFP at intracellular pH would be ∼3.9 (Akerboom et al. 2012). GCaMP6s is more sensitive but has a steep slope below 0.5 µM free Ca²⁺. GCaMP6f provides a more graded response up to 1 µM free Ca²⁺. GCaMP7b is significantly brighter at very low free Ca²⁺ concentrations. The exact physiological ranges of free Ca²⁺ in C. elegans neurons are not known and might differ depending on cell class and sub-cellular location. In practice, the advantages of GCaMP6s and GCaMP7b might be hampered by their expected stronger Ca²⁺-buffering capacities. Insert: Hill equation modified to estimate brightness levels and response curve: B: brightness; ε: extinction coefficient; ϕ: quantum yield; sat: Ca²⁺-saturated; apo: Ca²⁺-unbound; k_d: dissociation constant; n: Hill coefficient; [Ca²⁺]: free Ca²⁺ concentration. In vitro values for the parameters were taken from Dana et al. (2019). Note that they can vary with pH and temperature.

When choosing a GECI, the following interrelated biophysical properties should be considered.

1. Brightness: the usefulness of a GECI obviously depends on its detectability by the available imaging system. While the first two versions of GCaMP were orders of magnitudes dimmer than enhanced green fluorescent protein (EGFP) (Tallini et al. 2006), as from generation GCaMP5 the probes approximately match the brightness of EGFP in the Ca²⁺-bound form (Akerboom et al. 2012). Typically, methods papers report brightness with in vivo and in vitro levels. However, comparable across studies is when these values were measured in vitro. Brightness is expressed in the product between the sensor’s extinction coefficient (ε) and quantum efficiency (ϕ). It is important to pay attention to brightness levels in both the Ca²⁺-unbound (apo) and Ca²⁺-saturated states. Neurons at rest should be distinguishable from background levels and tissue-autofluorescence. FRET-based indicators typically have high brightness levels, close to their originating fluorophores. However, the relative changes in the fluorescence of the individual fluorophores upon FRET can be low, despite the ratio-signal providing a high dynamic range. If uncorrelated noise in the background is comparably high, this could render FRET-based indicators less useful in practice.

2. Dynamic range: the dynamic range, i.e., the relative signal change upon Ca²⁺-binding is critical. Note, that most novel GCaMP sensors improved dynamic range by lowering brightness in the apo state. This represents a significant trade-off, because in a neuron at rest, signals can be extremely low (10–40-fold when comparing with EGFP). Experimenters might be tempted to vastly overexpress the GECI risking cytotoxicity and Ca²⁺-buffering (see below).

3. Ca²⁺-binding affinity, expressed by the sensor’s dissociation constant (k_d), determines the sensitivity to low Ca²⁺-concentrations. High affinity comes at a trade-off of high Ca²⁺-buffering effects (discussed below) and slow Ca²⁺-unbinding rates, reducing the sensor’s temporal resolution.

4. Ca²⁺-binding cooperativity and response curve: GECIs have more than one Ca²⁺-binding site which cooperatively affect each other. This effect is expressed in the sensor’s hill constant. In combination with k_d, cooperativity determines the sensor’s response to free Ca²⁺. Ideally, the response curve is near-linear in the physiological range of free Ca²⁺. It can be calculated using the Hill equation (Figure 7C).

5. The off-rate (k_off or τ_{1/2, off}) determines how fast Ca²⁺ gets unbound from the sensor. This value is related to the affinity and is critical for the temporal resolution of the sensor.

The intracellular milieu exhibits a plethora of potential Ca²⁺-buffers making it difficult to estimate the relative contribution of an overexpressed GECI. Under the reasonable assumption that cytosolic free Ca²⁺ in C. elegans neurons ranges from 50 to 100 nM at rest to about above 1 µM when active, and that FPs are expressed in the µM range, like in heterologous cell culture systems (Cherkas et al. 2018), Ca²⁺-buffering by the probe is a conceivable issue (McMahon and Jackson 2018). Buffering has the deleterious effect of interfering with Ca²⁺-dependent signaling pathways. Therefore, buffering significantly can affect cell function, which was directly shown for ASH neurons in worms (Ferkey et al. 2007). Buffering dampens the peak response that can be measured in transient Ca²⁺-signals and prolongs the signal when Ca²⁺-levels decline, hence lowering temporal resolution (McMahon and Jackson 2018). Buffering is a function of both sensor concentration and its Ca²⁺-affinity. A direct way to test for Ca²⁺ buffering effects is to plot peak ΔF/F₀ responses, and on/off-rates versus F₀. A negative and positive correlation of these parameters respectively is a strong hint for buffering. To reduce buffering, we recommend to express only as much sensor as needed, or using sensors with lower affinity. Nuclear localized GCaMPs have been proven to work well in practice (Schrödel et al. 2013). We assume that Ca²⁺ can effectively diffuse through the nuclear pore complex. In some neurons tested side-by-side, nuclear- and cytosol- localized GCaMP delivered qualitatively similar results (Schrödel et al. 2013; Kato et al. 2015). In our experience, generating transgenic lines using pan-neuronally expressed GCaMP was more efficient in case of nuclear-localized versions, suggesting less cyto-toxicity (Schrödel et al. 2013). However, we never assessed this effect quantitatively and the underlying mechanism remains elusive.

Larsch et al. performed a systematic study, comparing various versions of cytosolic and nuclear GCaMP in AWA sensory neurons responding to the odorant diacetyl. Interestingly, odor-off responses were approximately an order of magnitudes slower than the theoretical values for any version of GCaMP (assuming an immediate off-step in Ca²⁺). These data indicate that Ca²⁺-dynamics in these neurons likely operate on longer timescales and that GCaMP kinetics are sufficiently fast to measure them. However, Ca²⁺-buffering also likely contributed to these slow dynamics: cytosolic GCaMP5a improved twofold compared to its older peer cytosolic GCaMP2.2; interestingly, nuclear GCaMP6f performed equally to cytosolic GCaMP5a, by these criteria (Larsch et al. 2015). These results suggest that the expected loss in temporal resolution due to the diffusion barrier represented by the nuclear compartment can be acceptable. However, as discussed below, some neurons exhibit Ca²⁺-dynamics restricted to their processes; such sub-cellular dynamics are obviously undetectable by nuclear sensors.

As a rule of thumb for generating new transgenic lines, we recommend wisely choosing between established and the newest GECIs on the market, and to evaluate the available choices by the above criteria. The modified Hill equation (Figure 7C) can assist you in estimating how the sensor might behave; ideally, it would show a near linear response curve from the nM to the µM range, has a fast k_off and is well-detectable in the neurons at rest. Because we do not know the exact physiological ranges of free Ca²⁺ for each worm neuron, the best sensor must be always determined empirically. Figure 7C also illustrates why it is recommendable to try at least two different sensors with different operating ranges. It is worth the effort to codon optimize the constructs and to introduce introns that boost expression. Next, we recommend generate multiple transgenes that express at different levels and then perform some preliminary experiments to compare them against each other, including the aforementioned test for buffering effects. In general, only express as much GECI as necessary. Most users will perform their Ca²⁺-imaging studies alongside with some behavioral experiments; therefore, imaging strains can be tested whether wild-type behavior is affected in them. Table 1 lists a selection of GECIs and their properties.

Table 1.

A list of GECIs useful for applications in C. elegans

GECI	Key properties	Notes and references
GCaMP6f	Fast, gradual response over physiological range of free Ca2+	Freely moving Ca2+ imaging in head motor neurons and nuclear localized version for whole-brain imaging (Kaplan et al. 2020)
GCaMP6s	Sensitive to low free Ca2+ concentrations	Used in many studies, including whole-brain imaging (Venkatachalam et al. 2016), also when localized to nucleus (Nguyen et al. 2016)
jGCaMP7b	High baseline fluorescence	Useful for imaging from subcellular locations (Dana et al. 2019)
GCaMP-X	Includes ApoCaM domain to prevent associating with intracellular CaM binding partners	Reduced cyto-toxicity in rodent neurons (Yang et al. 2018)
YC3.60	Ratiometric FRET indicator with good dynamic range	Broad application in C. elegans neurons (Nagai et al. 2004)
Tn-XXL	Troponin based indicator	Perhaps reduced cytotoxicity (Mank et al. 2008)
IP2.0	Sensitive to Ca2+ decreases	Validated in C. elegans sensory neurons (Hara-Kuge et al. 2018)
jRCaMP1 & jRGECO	Family of red shifted GECIs	Validated in C. elegans sensory neurons (Dana et al. 2016)
XCaMPs	Blue-, green-, yellow-, and red shifted GECIs	(Inoue et al. 2019)
Ribo-GCaMP, SomaGCaMP	Soma targeted GCaMPs	Used in pan-neuronal imaging (Chen et al. 2020; Shemesh et al. 2020)

The perspectives of voltage imaging

In recent years, various types of genetically encoded voltage indicators (GEVIs) have been developed (see, Pal and Tian 2020; Shen et al. 2020) for more comprehensive reviews. Voltage sensing domain (VSD)-based indicators [e.g., butterfly (Akemann et al. 2012) or ASAP3 (Villette et al. 2019)] couple the molecular motion of a VSD to a circular permutated FP or a FRET FP pair. Opsin-based GEVIs like Archon (Piatkevich et al. 2018) or QuasAr (Zou et al. 2014) exhibit a retinal chromophore that becomes protonated upon depolarization leading to increased fluorescence emission. FRET-to-opsin GEVIs introduce a FP FRET donor to an opsin; Vm increases are thereby reported via FP fluorescence decreases (Kannan et al. 2018). The chemigenetic GEVI Voltron operates by the same principle but the FRET donor requires a synthetic dye that needs to be complemented (Abdelfattah et al. 2019). Voltron improved some drawbacks of other GEVIs, i.e., low fluorescence intensities and small fluorescence changes in response to Vm fluctuations.

These drawbacks of GEVIs, however, represent a challenge to record the graded and small Vm fluctuations in C. elegans cells (Azimi Hashemi et al. 2019). Earlier studies successfully reported Vm recordings in thermosensory ADF neurons using (Kuhara et al. 2011) VSD-based mermaid (Tsutsui et al. 2008). VSD-based GEVIs of the butterfly family were successfully employed in AIY neurons (Shidara et al. 2013) using VSFP2.42 (Akemann et al. 2010) and AIA neurons (Lorenz Fenk & Mario de Bono, personal communication) using VSFP butterfly 1.2 (Akemann et al. 2012). Archon1 was validated in AVA neurons (Piatkevich et al. 2018). To date, Hashemi et al. provided the most comprehensive study comparing various GEVIs in worms. The GEVIs Arch(D95N), Archon, QuasAr, MacQ-mCitrine were successfully used in muscles; QuasAr reported excitatory and presumably inhibitory events in RIM interneurons (Azimi Hashemi et al. 2019). In conclusion, the application of GEVIs in C .elegans is still in its infancy but these reports are extremely promising. Given how GECIs transformed C. elegans neurobiology, the rapid development of GEVIs, will certainly enable exciting new avenues for future research.

Experimental setups

Microscopy:

The transparency and small axial dimensions make C. elegans highly amenable to in vivo fluorescence microscopy. Ca²⁺-imaging using GECIs is typically performed with high resolution 40x or 63x objectives, but high-quality data could be acquired with as low magnification as 5x (Larsch et al. 2015) (see below). Depending on the experimental paradigm, acquisition speeds typically range from 5 to 30 Hz.

If individual neurons are recorded, a standard epi-fluorescence microscope suffices but we recommend usage of bandpass filter sets to minimize autofluorescence and reliable control over excitation intensity. On the detector side, a sensitive EMCCD or a modern sCMOS camera is recommended with a dynamic range of at least 12 bit; the more sensitive the camera is, the more control over bleaching is in hand of the experimenter. However, camera gain should be used sensibly to achieve optimal SNR. It is important that acquired signals never saturate the detector; plan your experiments ahead accordingly when choosing light-intensity, exposure-time, and gain settings. Importantly, communication with the acquisition software must ensure regular acquisition time-intervals that can be recorded. If available, use the camera’s streaming mode, so that no emitted photon gets wasted.

Multi-neuron imaging from more than one focal plane sets higher requirements for the imaging setup. While it is possible to acquire neuronal population activity data from thousands of cortical neurons sparsely firing action potentials, even using wide-field excitation, such experiments can be quite challenging in worms. Its neurons are much smaller and densely packed into ganglia where adjacent somas, less than 1 µm apart, exhibit different continuous activity patterns, putting high demands on the optical resolution. C. elegans, however, is a relatively thin and low light-scattering sample compared to flies, fish, or mice. Therefore, point scanning 2-photon microscopy approaches do not necessarily provide any advantage in this case. Spinning disk confocal scanning microscopy was so far the method of choice in the vast majority of studies (Kato et al. 2015; Kotera et al. 2016; Nguyen et al. 2016; Venkatachalam et al. 2016). Here, images from a large excitation plane can be conveniently acquired through high N/A objective lenses at sufficient optical resolution onto a camera detector. The sample can be scanned in the z-dimension rapidly using a Piezo stage mounted either to the objective or sample holder. In order to achieve the sufficient volume acquisition rates (e.g., 4–5 vps, at least 15 × 2 µm spaced planes per volume) the system must be optimized for speed, i.e., (I) high sensitivity to allow low-exposure times and (II) synchronous operation of the z-stage with the camera in streaming-mode. This is typically hardly achieved with commercially available setups and requires sophisticated customizations. Alternative approaches have been used: wide field 2-photon imaging combines the advantages of multi-photon imaging with plane acquisition (Schrödel et al. 2013), but comes at high investment costs, requires expertise in maintenance, and restricts the lateral acquisition dimensions (∼70 × 70 µm vs >100 × 100 µm spinning disk). Light-field deconvolution microscopy (LFDM) (Prevedel et al. 2014) acquires images through a lens array allowing reconstruction of volumetric images from a single camera shot. The trade-offs of LFDM are reduced optical resolution, optical artifacts, need for high excitation power, and extensive computations for image reconstruction. Volumetric imaging via light-sheet microscopy (LSM) offers largely reduced phototoxicity, however, the perpendicular geometry of excitation-emission beam-paths precludes imaging worms in microfluidic devices or agarose-pads (see below). New variants of LSM, termed SCAPE2 (Voleti et al. 2019) and Snouty (https://andrewgyork.github.io/high_na_single_objective_lightsheet/), circumvent this problem by generating the light sheet through the detection objective. SCAPE2 reported high-quality Ca²⁺-imaging data from moving C. elegans; these LSM approaches potentially represent great advances over the spinning disk microscopy setups currently employed and perhaps will set the benchmarks in the following years.

Imaging in immobilized animals:

Imaging on agarose pads. Immobilization of C. elegans to microscopy stages offers the opportunity to manipulate animals and their environments in a highly controlled manner while recording neuronal activities. The first Ca²⁺-imaging studies adopted the protocol that was previously used for electrophysiology, and which is still frequently applied: here, worms are glued to an agarose surface and immersed in buffer using a perfusion chamber. Stimuli of soluble compounds are delivered by a microcontroller-actuated capillary moving towards and away from the animal’s nose (Hilliard et al. 2005; Suzuki et al. 2008). Similarly, mechanical stimuli can be applied to the cuticle at defined body locations (Suzuki et al. 2003; Kindt et al. 2007). Defined profiles of thermal stimuli can be established on agarose pads covering the animals with a glass-slide and upon coupling to temperature control elements (Kimura et al. 2004; Clark et al. 2006; Beverly et al. 2011; Kuhara et al. 2011). The preparation in the perfusion chamber further offers the opportunity to record in semi-restrained worms by gluing them only at the head; this enabled some of the first imaging studies of motoneurons (Haspel et al. 2010). In combination with a gas-flow delivery device (e.g., a transparent Y-shaped chamber), oxygen and carbon dioxide sensory responses in worms glued onto agarose pads can be recorded (Persson et al. 2009; Carrillo et al. 2013).

Imaging in microfluidic devices:

Microfluidics have brought applications in Biology (Sackmann et al. 2014). A recent wormbook study provides an in-depth overview for microfluidic applications in C. elegans, including Ca²⁺-imaging (San-Miguel and Lu 2013); therefore, we just summarize here some of the key applications relevant for imaging of neuronal activity. Microfluidic devices are manufactured from optically transparent silicon polymers (PDMS) in a photolithography procedure enabling the design of any desired 2D-pattern, or multi-layered  D-pattern of µm-scale fluidic channels. In these dimensions laminar flow prevails, enabling the experimenter with great control over the chemo-sensory environment. The dimensions of C. elegans body and its long-term viability in aqueous solution make it well suited to be studied in such devices. When placed in a worm-sized microfluidic channel, its head can be kept under a high magnification microscopy lens and imaged through a cover glass onto which the device is bonded, thus enabling the use of high N/A immersion lenses.

Water soluble stimuli can be delivered to the worm’s nose in the olfactory chip via controlling buffer-streams (Chronis et al. 2007); this device enables precise temporal control to switch stimuli at the sub-second scale (Kato et al. 2014). The olfactory chip was further advanced by incorporating a design to switch between 4 stimuli in arbitrary order (Rouse et al. 2018). In the widely used olfactory chip, stimulus and control-buffer flows are directed to or away from the nose by actuator flows; this configuration was meant to minimize pressure profiles in the device that are caused by valve-switches (Chronis et al. 2007). Every imaging experiment using the olfactory chip and its derivatives should never refrain from performing the adequate control experiment, simply switching buffer-to-buffer. Pressure profiles cannot be brought down to absolute zero in such devices and we do not know, which neurons in C. elegans remain sensitive to them.

Microfluidic chips were developed for the delivery of oxygen and carbon dioxide stimuli; the diffusion constants of these gases in PDMS are similar to water conditions, therefore they can rapidly diffuse to the worm through a thin PDMS membrane in a 2-layer device (Zimmer et al. 2009; Drexel et al. 2016).

Microfluidic devices are also well suited to deliver mechanical stimuli. Pressure controlled actuators were designed to study proprioception of enforced body postures (Wen et al. 2012), or responses to external mechanical stimuli of defined force delivered to defined locations at the cuticle (Cho et al. 2017; Nekimken et al. 2017).

Microfluidics can be also used to correlate neuronal activity with movements. Optimizing the channel dimension can lead to a treadmill situation, where worms generate undulatory waves without progressing; this approach enabled the first imaging study showing that AVA backward interneurons lock to reversal behavior (Chronis et al. 2007). In a similar approach, it is possible to fix the worm’s body while allowing for head movements, which revealed fundamental insights into the complex circuitry of head motor-neurons, and how sensory stimuli can get shaped by these movements (Hendricks et al. 2012; Shen et al. 2016; Liu et al. 2018a). McCormick et al. (2011) optimized a device to precisely position the head and to study the animals’ responses to dorsally or ventrally delivered stimuli.

Microfluids further enable researchers to perform high-throughput imaging experiments. Worms in a microfluidic arena (Albrecht and Bargmann 2011) can be effectively paralyzed with the ACh-agonist levamisole. Using low magnification lenses many worms at once could be imaged in parallel, enabling systematic studies of chemosensory-responses and their trial-to-trial variability (Larsch et al. 2013, 2015). Alternatively, worms can be loaded into microfluidic chips in an automatic and sequential manner, avoiding the trade-off of low magnification and drastically enhancing the throughput of high-resolution imaging experiments (Chung et al. 2008; Bazopoulou et al. 2017).

In summary, worm researchers are spoilt for choices in the configurations and experimental paradigms for immobilized Ca²⁺-imaging experiments. While imaging on agar pads offers the opportunity to better match the environment during imaging experiments to the worm’s cultivation conditions, microfluidic devices offer the advantage of better control and enhanced reproducibility of environmental and stimulus conditions. Note, that one caveat of microfluidic devices is that worms tend to enter a sleep-like quiescent state. The onset timing of this state is only partially predictable and occurs after minutes to hours in the device, depending on the degree of confinement (Gonzales et al. 2019). How microfluidic-induced sleep contributes to trial-to-trial variability observed in Ca²⁺-imaging studies, and whether it occurs in glued worms as well, is not known.

Ca²⁺-imaging in freely moving animals:

Several approaches have been developed to enable Ca²⁺-imaging in freely behaving animals. Low-magnification imaging in the microfluidic arena mentioned above was a straightforward way to capture the activity of neurons from animals freely navigating in these devices (Larsch et al. 2013). Most studies, however, use online-tracking methods to keep the region of interest (ROI) in the center of a high magnification lens. This can be achieved either by hand (Busch et al. 2012), with various software applications (Ben Arous et al. 2010; Nguyen et al. 2016; Tanimoto et al. 2016; Venkatachalam et al. 2016) or via a hardware loop using a quadrant photomultiplier (PMT) (Faumont et al. 2011). The advantage of the hardware system is that it operates nearly in real-time; the trade-off of this system is that it requires a high SNR signal, meaning that a significant proportion of the emitted fluorescence light must be diverted to the PMT and is therefore lost. In combination with a dual emission beam-path, this loss of signal can be restricted to a reference channel avoiding any loss of the GECI signal (Kato et al. 2015).

Movement artifacts, i.e., tracking errors and out-of-focus movements, are a major concern in these experiments. Therefore, we highly recommend to either use ratiometric GECIs or to co-record a reference signal emitted at a different wavelength and detected through a dual emission beam-path. Artifacts can be corrected by calculating the signal-ratio or other computational methods like independent component analysis (Hallinen et al. 2021).

Typically, behavior is recorded through a separate camera mounted on the opposite side, ideally capturing the entire animal and at lower magnification with darkfield or infrared conditions (Faumont et al. 2011; Nguyen et al. 2017). Such behavioral co-recordings enable correlative studies relating neuronal activity with detailed postural dynamics (Hums et al. 2016; Kaplan et al. 2020; Hallinen et al. 2021).

Using these techniques, animals can freely roam in microfluidic arenas (Hallinen et al. 2021). Alternatively, they can crawl on agarose pads, i.e., conditions nearly identical to cultivation conditions in agarose dishes. Stimuli can be provided as odorant gradients (Liu et al. 2018a). Moreover, the agarose pad can be sealed by a glass surface in a customized device that allows, via a small headspace, gaseous stimuli like oxygen and carbon dioxide to be delivered (Kato et al. 2015; Hums et al. 2016). In both cases, imaging is performed via long-distance lenses. Other approaches use a sandwich configuration, with a thin glass slide placed directly on top of the agarose pad (Nguyen et al. 2016; Venkatachalam et al. 2016; Kaplan et al. 2020). This provides the advantage of permitting high N/A oil or water immersion lenses drastically improving image quality. Under these conditions, thermal gradients can be coupled to the arena to serve as sensory stimuli (Clark et al. 2007; Nguyen et al. 2016).

Whole-brain Ca²⁺-imaging:

Using various combinations of the techniques described in this section, it is possible now to record nervous-system wide imaging in real-time and at single-cell resolution putting C. elegans onto the stage as a prime model organism for systems neurobiology. As mentioned above, dense packing of small somas represents a challenge to segmenting the resulting images when GECIs are expressed from pan-neuronal promoters. Venkatachalam et al. (2016) used cytosolic GCaMP in combination with nuclear localized TagRFP to segment ROIs for subsequent GCaMP measurements. Other studies used nuclear localized GCaMP alone (Schrödel et al. 2013) or in combination with red markers (Kotera et al. 2016; Nguyen et al. 2016). Recently, cell-body-targeted versions of GCaMP were introduced that combine the advantages of both cytoplasmic and nuclear localized GECIs (Chen et al. 2020; Shemesh et al. 2020); one study reported successful application in C. elegans (Chen et al. 2020).

Using microfluidic devices, it is possible to position head and tail in the imaging arena to record simultaneously from all major head-tail ganglia, retrovesicular ganglion, and parts of the ventral cord (Kaplan et al. 2020). Using wide-field imaging in combination with deconvolution it was even possible to record the entire nervous system, albeit by the expense of spatial resolution (Kaplan et al. 2020).

In combination with tracking stages and fast volumetric acquisition as described above, whole-brain imaging of head ganglia and male tail ganglia in freely moving worms was performed (Nguyen et al. 2016; Venkatachalam et al. 2016; Susoy et al. 2020; Hallinen et al. 2021), making now possible to relate network activity to instantaneous unrestrained behavior.

Whole-brain imaging in immobilized worms combines all advantages of controlled environments and available sensory stimulation paradigms, while delivering high SNR neuronal activity time-series. To enable sufficient volume acquisition rates and to avoid image blurr due to motion, whole-brain imaging in feely moving worms must be performed with very high framerates, thus limiting camera exposure times. Moreover, the nonrigidly deforming tissues of moving worms represent a major challenge to extract clean neuronal activity time series from such low SNR images. The field is currently very active in overcoming these challenges by developing sophisticated image processing pipelines (Tokunaga et al. 2014; Toyoshima et al. 2016; Nguyen et al. 2017). Notably, Nguyen et al. (2017) developed a promising approach that refrains from frame-to-frame tracking, thus preventing error propagation, via a time-independent registration algorithm. The former section of this study describes in-depth modern approaches to extract cell-identities from volumetric whole-brain imaging data.

Important considerations for designing and interpreting your Ca²⁺-imaging experiments

Subcellular Ca²⁺-dynamics:

Several studies reported Ca²⁺-signals in various interneurons that were restricted largely to the neurites, e.g., in AIY, AIZ, and RIS (Chalasani et al. 2007; Larsch et al. 2013; Li et al. 2014; Costa et al. 2019). Moreover, in RIA neurons, different domains of the process compute information separately, partially even in an anti-phasic manner: while the proximal domain of the RIA neurite mainly represents sensory inputs, two distal sub-domains (nrD and nrV) receive anti-phasic corollary discharge signals from motorneurons (Hendricks et al. 2012; Liu et al. 2018a).

One major lesson from these studies is that whole-brain imaging using nuclear- or soma-localized GECIs, obviously cannot capture the full spectrum of neuronal activities that can be found in the worm brain. Moreover, in every neuronal imaging study, researchers should also pay attention to possible signals in the neurites, which is more challenging for the imaging analysis methods, especially in freely crawling animals.

But what is the correct interpretation of local Ca²⁺-signals? These observations seem to be in contradiction with electrophysiology results leading to the proposal that C. elegans neurons should be isopotential. In the case of RIA, antiphasic signals speak strongly in favor of compartmentalized Ca²⁺-signaling domains. In fact, they receive compartmentalized cholinergic input from motorneurons via muscarinic acetylcholine (Ach) receptors (Hendricks et al. 2012; Liu et al. 2018a). These receptors signal via G-proteins, suggesting that Ca²⁺-serves as a second messenger signal without associated strong local fluctuations in Vm. This model, however, must not be generalized to all neurons where Ca²⁺ signals seem restricted to neurites. Figure 8 illustrates that due to the thin diameters of neuronal processes in worms Ca²⁺-entry through ion channels causes large changes in local Ca²⁺-concentrations and Ca²⁺-pumps can execute their function rapidly to restore resting Ca²⁺. In the cell body, due to the larger volume, the effect on global Ca²⁺-concentrations is much smaller. In addition, buffering by the cytosol and the overexpressed GECI further diminish and convolute transient changes in cytosolic free Ca²⁺. The local concentration changes at the cell membrane, however, can be transiently as strong as in the neurite, but would be detected as tiny changes in the integrated intensity of a globally expressed GECI. In this case, we recommend to be cautious with the conclusion about compartmentalized Ca²⁺-domains, without further experimental validation. Evidence for this interpretation comes from Liu et al. co-measuring Vm, neuritic and somatic Ca²⁺ in AWA. While neuritic Ca²⁺-signals resembled the shape of AWA action potentials, in the soma a largely convoluted signal was measured (Liu et al. 2018b).

Figure 8.

Illustration of Ca²⁺-dynamics in neurite and soma. Shown is a simplifying schematic of a C. elegans neuron with an axo-dendritic neurite and soma (this morphology is shared e.g., among many head interneurons). In small diameter neurites, Ca²⁺-entry through ion channels causes large transient amplitudes in local Ca²⁺-concentrations, which can be rapidly restored to resting levels via Ca²⁺-pumps. In the cell body the same amount of Ca²⁺-entering causes a strong local- but small change in global Ca²⁺ concentrations due to the much larger volume. Other effects, like Ca²⁺-buffering of the cytosol via the bulk of Ca²⁺ binding proteins can further dampen and convolve the measurable Ca²⁺-signals.

Spontaneous neuronal dynamics and brain states:

Worms switch their behavioral states, like forward-backward crawling (Roberts et al. 2016), roaming-dwelling (Flavell et al. 2013), or wake-to-sleep (Nichols et al. 2017) in a seemingly stochastic and spontaneous manner, i.e., without the need of external triggers at the onset of each transitions. For an example, see the rigorous activity of many neurons in the absence of stimulation in Figure 3. In particular, previous work showed that forward-backward crawling is represented by large populations of interneurons and motorneurons, the activity of which reliably locks to these transitions, an observation that can be made even in immobilized conditions (Kato et al. 2015). Surprisingly, this feature was found in primary sensory interneurons previously thought to exclusively represent sensory information: recording for example AIY, RIB, or AIB neurons in freely moving worms in the absence of acute sensory stimuli revealed that they are strongly modulated by the animals' instantaneous behavior and become reliable activated by forward or reversal states (Li et al. 2014; Luo et al. 2014; Kato et al. 2015; Laurent et al. 2015), the possible functions of these global behavioral representations were discussed recently (Kaplan et al. 2018; Kaplan and Zimmer 2020). These findings have strong implications in how to design and interpret Ca²⁺ imaging experiments. Any neuronal activity measurement, whether in immobilized or freely moving worms, occurs in the context of diverse network activities that represent ongoing behavior and brain states (Kato et al. 2015; Nichols et al. 2017; Hallinen et al. 2021). For example, any of the neurons that are locked to behavioral states, e.g., reversals, will show a stimulus locked trial-averaged signal when worms are presented with stimuli that trigger transitions between forward and backward crawling. However, such a result cannot distinguish whether the neuron under study is directly modulated by the sensory input, the current motor-state, or both. We previously suggested an experimental design that helps interpreting interneuron imaging results in this respect (Kaplan et al. 2018). First, it is important to record the baseline of each neuron of interest in the absence of any stimulation for a sufficient amount of time. Given that in immobilized conditions forward, like backward, command states or sleep can last up to a minute or more (Kato et al. 2015; Nichols et al. 2017) it is needed to record baseline activity levels for several minutes. In case neurons exhibit spontaneous activity or bi-stability, the effect of stimulation on features like signal amplitude, -frequency, -width and -probability at stimulus onset can be calculated and compared to baseline values. See Kaplan et al. (2018) for a recommended analysis pipeline and detailed discussion.