Connectomes across development reveal principles of brain maturation in C. elegans

Abstract

From birth to adulthood, an animal’s nervous system changes as its body grows and its behaviours mature. However, the extent of circuit remodeling across the connectome is poorly understood. Here, we used serial-section electron microscopy to reconstruct the brain of eight isogenic C. elegans individuals at different ages to learn how an entire wiring diagram changes with maturation. We found that the overall shape of the nervous system is preserved from birth to adulthood, establishing a constant scaffold upon which synaptic change is built. We observed substantial connectivity differences among individuals that make each brain partly unique. We also observed developmental synaptic changes that are consistent between animals but different among neurons, altering the strengths of existing connections and creating additional connections. Collective synaptic changes alter information processing of the brain. Across maturation, the decision-making circuitry is maintained whereas sensory and motor pathways are substantially remodelled, and the brain becomes progressively more modular and feedforward. These synaptic changes reveal principles by which maturation shapes brain and behavior across development.

Introduction

The developing nervous system faces multiple challenges. Amid an animal’s changing anatomy and fluctuating environment, some circuits must maintain a robust output, such as loco-motion ^1–4. New circuits need to be constructed in order to support new functions, such as reproduction ^5–7. Moreover, to adapt and learn, the nervous system must make appropriate changes in existing circuits upon exposure to external cues ⁸. Neural systems employ a variety of adaptive mechanisms to overcome these challenges. In the Drosophila nerve cord, synaptic density of mechanosensory neurons scales to body size from first to third instar larvae ⁴. In the spinal cord of the zebrafish larva, descending neurons lay down tracks chronologically, coinciding with the maturation of swimming behaviors ⁷. In the mouse visual circuit, postnatal synaptic remodeling is shaped by intrinsic activity as well as visual stimuli ⁹. The prevalence of anatomical changes in the nervous system must accommodate both growth and experience.

Anatomical changes occur at many levels, from individual synapses to global organization of brain networks ¹⁰. An assortment of genetic and cellular factors have been found to affect morphological and functional maturation of synapses ¹¹,¹². Synaptic changes are also likely to be coordinated across developing circuits, giving rise to system-level modifications. However, developmental principles that describe the synaptic changes that shape the adult brain are unknown.

Interrogating whole-brain maturation at synapse resolution is difficult. High-resolution reconstruction is needed to capture structural changes at individual synapses ¹³. These methods must be applied to an entire brain, and to brains at different developmental timepoints. Moreover, multiple animals need to be be analyzed to assess structural and behavioral heterogeneity. Electron microscopy (EM) allows reconstruction of neural circuits with synapse resolution ^14–20, but low throughput makes it difficult to compare whole brain samples and comprehensively quantify plasticity. EM has been applied to assess wiring differences between species ²¹, sexes ²², genotypes ²³, and ages ⁴,²⁴. But previous studies mapped partial circuits or few samples.

The original C. elegans connectome was compiled from the EM reconstruction of partially overlapping regions of at least four adults and an L4 larva ^25,26. A revisit of the C. elegans connectome expanded this wiring diagram by re-annotation of original EM micrographs and filled remaining gaps by interpolation ²². Such compilations make it difficult to assess plasticity, variability, or correlations between individuals.

Here, we leveraged advances in the automation and throughput of EM reconstruction to study of the brain of C. elegans - its circumpharyngeal nerve ring and ventral ganglion - across development. We have fully reconstructed the brains of eight isogenic hermaphroditic individuals at different ages from birth to adulthood. These reconstructions provide quantitative assessments for the length, shape, and position of every neuron and fiber in the nerve ring, as well as of every physical contact and chemical synapse between neurons, glia, and muscles. Our quantitative comparisons of these developmental connectomes have revealed several organizing principles by which synaptic growth and remodeling shape the mind of the developing worm.

Results

EM reconstruction of eight C. elegans brains from birth to adulthood

We leveraged advances in ultra-structural preservation, serial ultra-thin sectioning, and semi-automated imaging ^27–29 to reconstruct the connectivity and shape of eight individual isogenic hermaphroditic brains of C. elegans (N2) at various post-embryonic stages (Fig. 1a, Fig. S1, Video 1-2, see Methods). The brain, consisting of the nerve ring and ventral ganglion, includes 162 of the total 218 neurons at birth (L1), and 180 of the total 300 neurons in adulthood (Table S1; excluding CANL/R throughout development and HSNL/R until adulthood) ^25,30. It also contains 10 glia and synaptic sites of 32 muscles at all stages. Using previous identifiers, we named every cell across different EM volumes based on their unique neurite morphology and position ²⁵. Each neuron was classified as either being sensory, inter, motor, or modulatory (Table S1, Video 2, see Methods).

Figure 1. The developing brain maintains topology.

a. Developmental timeline of eight reconstructed brains, with topological models shown at three stages. The models include all cells contained in the neuropil, colored by cell types. b. Wiring diagrams for all datasets. Each circle represents a cell. Each line represents a connection with at least one chemical synapse between two cells. Line width indicates synapse number. The vertical axis denotes signal flow from sensory perception (top) to motor actuation (bottom); the horizontal axis denotes connectivity similarity where neurons with similar partners are positioned nearby each other ²⁶. Signal flow and connectivity similarity are based on the accumulated connections from all datasets. c. A representative EM micrograph of the neuropil (from dataset 3). Classical chemical synapses are characterized by a pool of clear synaptic vesicles (red arrows) surrounding an active zone (red arrowhead). Chemical synapses of modulatory neurons are characterized by mostly dense core vesicles (orange arrows) distant from the active zone (orange arrowhead). Postsynaptic cells are marked by asterisks. d. The summed length of all neurites in the brain exhibits linear increase from birth to adulthood. Each data point represents the total neurite length from one dataset. e. Physical contact between neurites at birth (persistent physical contacts) accounts for nearly all of the contact area at every developmental stage. f. Total synapse numbers in the brain exhibits a 6-fold increase from birth to adulthood. g. Synapse density (the total number of synapses divided by the total neurite length) is maintained after an initial increase.

In each EM volume, every neuron, glia, and muscle were volumetrically segmented and annotated for chemical synapses to generate a complete connectome of the brain (Fig. 1b, Fig. S2, Video 2, see Methods). These reconstructions include classical synapses with mostly clear vesicles and synapses of modulatory neurons with mostly dense core vesicles (Fig. 1c, see Methods). We plotted the wiring diagrams conforming to the direction of information flow from sensory perception (Fig. 1b top layer) to motor actuation (Fig. 1b bottom layer). All connectomes are hosted on an interactive web-based platform at http://nemanode.org/. These datasets allowed for examination of synaptic connectivity in the context of topology, including the shape and size of each neuron as well as the proximity and contact between each neurite.

Uniform neurite growth maintains brain topology

Our volumetric reconstructions revealed striking similarities of brain topology between developmental stages. The shape and relative position of every neurite and cell body in the brain was largely established by birth (Fig. S3a). From birth to adulthood, the total length of neurites underwent a 5-fold increase (Fig. 1d), in proportion to the 5-fold increase in body length (∼250µm to ∼1150µm). Neurites grew proportionally (Fig. S3b), maintaining most physical contact between cells at birth across maturation (Fig. 1e). Only three neuron classes (RIM, ADE, and SAA) had changes to their primary branching pattern, each growing a new major branch after birth (Fig. S4). Thus, the brain grows uniformly in size without substantially changing the shape or relative position of neurites, maintaining its overall topology.

In parallel to neurite growth, addition of synapses was extensive from birth to adulthood. The total number of chemical synapses increased 6-fold, from ∼1300 at birth to ∼8000 in adults (Fig. 1f). We found no evidence for systematic synapse elimination. Presynaptic terminals appear as en passant boutons, most often apposing the main neurite of a postsynaptic cell. Small spine-like protrusions ^25,31 were postsynaptic at only ∼17% of synapses in the adult connectome (Fig. S3c). From birth to adulthood, the number of spine-like protrusions increased 5-fold (Fig. S3d), and the proportion of spine-like protrusions apposing presynaptic terminals increased 2-fold (Fig. S3e).

Synapse number increased in proportion to neurite length, maintaining a stable synapse density. However, during the L1 stage, the increase of total synapse number slightly outpaced that of neurite length, leading to increased synapse density (Fig. 1g). This increase coincided with an increasing left-right symmetry in connectivity (Fig. S3f, S3g). In the adult brain, ∼90% of neurons exist as left-right pairs that mirror one another in position, morphology, as well as connectivity. However, some of these neurons exhibited asymmetry in left-right connectivity at birth (Fig. S3f, S3g). The simplest interpretation of this early asymmetry is incompleteness. C. elegans hatches before its brain connectivity has been made symmetric, a process which continues by synapse addition during the first larval stage.

Non-uniform synapse addition reshapes the connectome

From birth to adulthood, addition of synapses both creates new connections and strengthens existing connections. Here, a connection is defined as a pair of cells connected by one or more chemical synapses (Fig. 2a). The 204 cells of the brain were interconnected by ∼1300 total synapses distributed among ∼800 connections at birth (Fig. 2b). Over maturation, addition of synapses strengthened nearly all existing connections. Approximately 4500 synapses were added to connections that were present at birth, such that the mean synapse number per connection increased 4.6-fold, from 1.7 synapses per connection at birth to 6.9 by adulthood (Fig. 2c). In addition, many new connections formed. Approximately 1200 synapses formed between previously non-connected neurons resulting in a 2.4-fold increase in total number of connections between cells present at birth (Fig. 2b).

Figure 2. Non-uniform synapse addition reshapes the connectome

a. Schematic of a connection. Each connection consist of at least one synapse between two cells. b. The total number of connections in the brain exhibits a 2.4-fold increase. c. The mean number of synapses per connection existing from birth exhibits a 3.9-fold increase. d. The probability of a new connection (a connection that appears in datasets 7 and 8 but is absent in datasets 1 and 2) that form at physical contacts existing from birth. This probability increases with the total contact area between two cells at birth. e. Top: neurons with higher number of connections at birth (dataset 1) are more likely to receive new synapses at existing input connections by adulthood (averaging datasets 7 and 8). Bottom: no positive correlation is observed at existing output connections. Each data point represents one cell. Significance is calculated using Spearman’s rank correlation. f. Top: neurons with higher number of connections at birth (dataset 1) are more likely to establish new input connections by adulthood (averaging datasets 7 and 8). Bottom: no correlation is observed at new output connections (bottom). Each data point represents one cell. Significance is calculated using Spearman’s rank correlation. g. Top: each data point represents the mean coefficient of variation (CV) in the number of synapses for different sets of connections. The CV of output connections from the same cell is maintained. The CV of input connections to the same cell increases over time, at the same rate as connections to and from different cells. Bottom: the difference between the mean CV for output and input connections relative to connections between different cells grows over time. *** p < 10⁻⁴, Spearman’s rank correlation coefficient.

Synapse addition did not occur uniformly across the brain. We found preferential synapse addition in multiple contexts.

First, we found that new connections were more likely to form at existing physical contacts between neighboring neurons with large contact areas (Fig. 2d). The physical contacts formed at birth appear to create a constant scaffold within which network formation unfolds.

Synapse addition was also not uniform between neurons. At birth, it was already evident that some neurons had far more connections than others (Fig. S5a). Neurons with more connections at birth disproportionately strengthened their existing connections over time (Fig. 2e). Interestingly, this disproportionate strengthening only occurred at input connections (Fig. 2e). Neurons with more connections at birth also disproportionately added new input connections in comparison to output connections (Fig. 2f). Thus, maturation focuses the flow of information onto the most highly-connected neurons at birth.

We found that synapse addition to existing connections also changes the relative strengths of a neuron’s inputs but not its outputs (Fig. 2g). We found no correlation in the strengthening of existing input connections to each cell from different presynaptic partners (Fig. S5b), leading to a divergence in their relative strengths (Fig. 2g). However, we observed that strengthening of the existing output connections from each cell were correlated (Fig. S5b), maintaining their relative strengths (Fig. 2g). Thus, each cell regulates the strengthening of its own synaptic outputs but does not dictate the relative strengthening of its inputs.

Isogenic individuals have both stereotyped and prevalent variable connections

We mapped the change in synapse number for each connection across all developmental stages. Thus, we were able to classify each connection as either stable, developmentally dynamic, or variable (Fig. 3a, Fig. S6). Stable connections were present from birth to adulthood and maintained their relative strength in proportion to one another. Developmentally dynamic connections significantly increased or decreased their relative strength in a stereotyped manner, sometimes even forming new connections or eliminating existing connections at specific life stages. Variable connections exhibited no consistent trend in their changing synapse numbers, and were not present in every animal.

Figure 3. Isogenic individuals have both stereotyped and prevalent variable connections.

a. A sensory circuit across maturation. Left: L1 (dataset 2), center: L3 (dataset 6), right: adult (dataset 8). Colour-coded lines represent stable (black), developmentally dynamic (blue), and variable (grey) connections. Line width represents synapse number. Cells are coloured by type. b. The total number of stable, developmentally dynamic, and variable connections in each dataset. c. The total number of synapses that constitute stable, developmentally dynamic and variable connections in each dataset.

In the adult connectome, stable and variable connections each represented ∼43% of the total number of connections, whereas developmentally dynamic connections represented ∼14% (Fig. 3b). Stable connections contained more synapses than variable ones (6.6±5.8 synapses versus 1.4±1.0 synapses, respectively, in adult), and thus constituted a large proportion (∼72%) of total synapses (Fig. 3b). Nonetheless, variable connections were surprisingly common. The number of variable connections in the adult (∼800) is similar to the number of stable connections (∼800). The number of variable synapses in the adult (∼1100) is even greater than that of developmentally dynamic synapses (∼800). Not all variable connections were weak (Fig. S7a). When connections between cell pairs with less than 4 synapses were excluded, variable connections still constituted ∼12% of all connections (Fig. S7b). Thus, variable connections make up a substantial proportion of the C. elegans connectome.

Variable connections are not uniformly distributed among cell types

To visualize the distribution of different classes of connections, we separately plotted their occurrences in the wiring diagram (Fig. 4a). Stable and developmentally dynamic connections represent the portion of the connectivity that is shared across animals. Variable connections represent the portion that is unique to each animal.

Figure 4. Non-uniform distribution of variable and developmentally dynamic synapses.

a. Wiring diagrams for variable, stable, and developmentally dynamic connections. Each line represents a connection observed in at least one dataset. Line width indicates the largest number of synapses observed for a connection across datasets. Each circle represents a cell. Cell coordinates are represented as in Fig. 1b. b. Comparison of the proportion of variable and non-variable connections for each cell type in the adult brain. Non-variable connections include stable and developmentally changing connections. Cell types with significantly higher or lower proportions of variable connections are indicated, ** p < 0.01, *** p < 0.001, Mann–Whitney U test, FDR adjusted using Benjamini–Hochberg correction. c. Wiring diagram showing non-variable connections between different cell types. Line width indicates the number of connections. Line color indicates the proportion of developmentally dynamic connections. Lines with significantly different proportions of developmentally dynamic connections are indicated, * p < 0.05, *** p < 0.001, two-tailed Z-test, FDR adjusted using Benjamini–Hochberg correction.

We quantified the proportion of variable connections in the inputs and outputs of each cell type (Fig. 4b). We found that modulatory neurons had significantly higher amounts of variability in their output connections than other cell types, whereas motor neurons had significantly less (Fig. 4b upper panel). Consistent with the lowest variability in motor neuron output, muscles exhibited the lowest variability in their inputs (Fig. 4b lower panel).

The non-uniform distribution of variable connections was still evident when weak connections were excluded (Fig. S7b). The low variability of connections from motor neurons to muscles could not be simply explained by saturation of their physical contacts by synapses (Fig. S7c). We also considered that neurons with more synapses may exhibit higher number of random developmental or annotation errors. However, the proportion of variable connections did not scale with the number of synapses (Fig. S7d-S7g). Rather, the likelihood of a neuron to generate variable connectivity is likely a property of its cell type. The high stereotypy of synapses from motor neurons to muscles may reflect a requirement for high fidelity in circuits for motor execution. Modulatory neurons, which can secrete monoamines and neuropeptides by volume-release, may have the weakest requirement for precise spatial positions of synaptic output because they exert long-range effects.

Interneuron connections are stable during maturation

Excluding variable connections allows us to properly assess developmental connectivity changes. We found that developmentally dynamic connections were not uniformly distributed among cell types or circuit layers (Fig. 4c). Connections between interneurons, and from interneurons to motor neurons had disproportionately more stable connections than developmentally dynamic connections (Fig. 4c). All other connections, between and from sensory, modulatory, or motor neurons, had many developmentally dynamic connections. Developmentally dynamic connections were particularly prevalent from motor neurons to muscles. Each motor neuron progressively recruited more muscles in a stereotypic pattern (Fig. S6). The abundant but high stereotypy of this developmental connectivity change means that motor neurons exhibit the lowest proportion of variable connections (Fig. 4b upper panel). Developmentally dynamic connections were also prevalent between many sensory neurons, and from sensory neurons to interneurons and motor neurons (Fig. 4c, Fig. S6).

These findings show that maturation changes how multisensory information is integrated and represented before it is relayed to downstream neurons. Maturation also changes motor execution. However, the layout of interneuron circuits, the core decision-making architecture of the brain, is largely stable from birth to adulthood.

Increase in both feedforward signal flow and modularity across maturation

With connectomes of complete brains across maturation, we were able to ask how the total set of synaptic changes leads to collective changes in information processing.

First, we examined how synaptic changes affect information flow in the brain. The directionality of signal flow between cells can be viewed as either feedforward, feedback, or recurrent (Fig. 5a). We classified connections that constitute synapses from the sensory to motor layer as feedforward, connections from the motor to sensory layer as feedback, and connections between neurons of the same type as recurrent. Among stable connections, synapse addition strengthened existing feed-forward connections more than feedback or recurrent connections (Fig. 5b). The addition of developmentally dynamic connections also preferentially increased feedforward signal flow (Fig. 5c). In contrast, those developmentally dynamic connections that weakened across maturation tended to be feedback and recurrent. Taken together, these changes gradually increase the proportion of feedforward synapses (Fig. 5d). Thus, one global pattern of brain maturation augments signal flow from sensation to action, making the brain more reflexive (and less reflective) with age.

Figure 5. Increase in both feedforward signal flow and modularity across maturation.

a. Schematic of feedforward, feedback, and recurrent connections defined by cell types. b. Number of synapses added to stable connections relative to the number of synapses at birth (dataset 1). Stable feedforward connections are strengthened more than stable feedback and recurrent connections. * p < 0.05, *** p < 0.001, Mann–Whitney U test, FDR adjusted using Benjamini–Hochberg correction. c. Proportions of feedforward, feedback, and recurrent connections for stable and developmentally dynamic connections. ** p < 0.01, *** p < 0.001, two-tailed Z-test of the proportion of feedforward connections, FDR adjusted using Benjamini–Hochberg correction. d. Proportions of the total number of synapses in feedforward, feedback, and recurrent connections. * p < 0.05, ** p < 0.01, *** p < 0.001, Spearman’s rank correlation. e. Number of cells in each module across maturation, determined by weighted stochastic blockmodeling. Modules connected by a line share significant number of neurons (see Table S2 for cell membership of each module). f. Wiring diagram for the adult connectome, with each cell colored by its assigned module. Cell coordinates are represented as in Fig. 1b. g. 3D model of the adult brain, with each cell colored by its assigned module.

Next, we asked how changes in connections affect the community structure of the brain. We used weighted stochastic blockmodeling (WSBM) to group neurons of similar connectivity into distinct modules ³². We found that the wiring diagram becomes more modular across maturation, increasing from two modules at birth to six modules in adults (Fig. 5e, Fig. S8a, Table S2). A similar increase was obtained with a generative evaluation framework, an independent estimator of modularity (Fig. S8b, see Methods). The increase in modularity can be mostly attributed to developmentally dynamic connections, which only represent 12% of total synapses (Table S3). Variable connections, which are not uniformly distributed among cell types, also contributed to module segregation (Table S3).

Increased modularity produces congregations of cells and circuits with functional specialization. At birth, sensory neurons and interneurons that relay and integrate sensory information were clustered into one module. By adulthood, labial sensory neurons (“anterior sensory”), amphid sensory neurons involved in taxis behaviors (“posterior sensory”), and remaining sensory neurons and the majority of interneurons (“medial sensory/interneuron”), became separate modules (Fig. 5e). At birth, head motor neurons and premotor interneurons that command body movements were clustered into the same module. In adult, they belonged to separate modules (“head movement” and “body movement”) (Fig. 5e, Table S2). Importantly, functional modules are created from closely connected neurons in the wiring diagram (Fig. 5f) as well as physically proximate neurons, reminiscent of distinct brain lobes (Fig. 5g).

Discussion

To learn if any principles emerge by studying the synaptic level structure of an animal’s nervous system across developmental stages, we analysed eight isogenic C. elegans beginning with the early larva and ending with the adult. While it took nearly a decade to analyze the first worm connectome ²⁵, the advent of automated sectioning, and both computer-assisted image acquisition and analysis greatly sped up the process allowing our complete brain reconstructions of many animals in far less time.

We found several general features that remained largely unchanged from the earliest larva to the adult. For example, the shape of the nervous system, its topology, and even the over-all three-dimensional shape of individual neurons were surprisingly stable throughout larval development. In contrast, the volumetric size of the nervous system, and the neurons that comprised it, enlarged about 6-fold. However, the wiring diagrams showed that developmental changes in the nervous system were not simply explained by enlargement of existing structures. Most significantly, there was a 5-fold increase in the number of synapses between connected neurons. These synaptic changes were not distributed uniformly through the network. Rather, they appeared to be organized by several principles that profoundly shape how the brain’s network changes with maturation.

The principles that we uncovered are illustrated in Fig. 6. At one level, that applies to every neuron in the brain, we observed patterns of synaptic remodelling that alter the number and strength of all connections. At a second level we observed synaptic remodelling that differed between cell types, (i.e., sensory neurons, modulatory neurons, interneurons, and motor neurons). At the third level, we observed network changes that altered the directionality of information flow and the segregation of information processing throughout the brain. It is likely that these many levels of synaptic remodelling (listed below) explain the ontogenetic basis of adult worm behavior:

Figure 6. Principles of connectivity changes across maturation.

Left: schematic of brain-wide synaptic changes from birth to adulthood. Right: emerging principles at the level of topology, individual neurons, neuron types, and the network that describe synaptic changes.

Large contacts predict new connections

Because the overall topology of the brain is constant, the physical contacts between neurites from birth to adulthood are nearly invariant. Nearly all the new synapses appear at sites where these physical contacts already exist, both adding synapses to connections between neurons and creating new connections between neurons. The larger the physical contact, the greater the probability of a new connection. Therefore, topology at birth creates the scaffold upon which adult connectivity is built.

More inputs to well-connected neurons

When we assayed developmental synapse addition among all neurons, we found that the cells with the largest numbers of connections at early stages received disproportionately many new synapses, both strengthening existing input connections and creating new input connections. In contrast, these neurons saw less synapse addition to their outputs. Thus, well-connected neurons became better integrators of information, but not broader communicators of that information.

Changing selectivity of a neuron’s inputs but not outputs

We also found a pattern in how synapses are selectively added to change the relative strengths of existing connections. The strength (synapse number) of input connections from different neurons that innervate the same neuron tend to become more heterogeneous. In contrast, the postsynaptic outputs of a neuron maintain their relative strengths. Neurons thus became more strongly driven by a subset of their presynaptic partners but distribute that information uniformly among their postsynaptic partners.

Prevalent variability in the connectomes between animals

Each animal has connections between neurons that were not found in other animals. These variable connections between neurons tend to be mediated by small numbers of synapses. Nonetheless, these variable connections represent almost half of all connected neuron pairs in the mature brain. This variability is most prominent among the modulatory neurons and least prominent among motor neurons.

Interneuron circuits are stable

We discovered remarkable stability in the wiring between interneurons that may constitute the core decision-making architecture of the brain. In contrast we found extensive developmental wiring changes among other cell types.

Feedforward bias increases

At the level of the entire network, we discovered a change in the directionality of information flow. Synaptogenesis over development preferentially creates new connections and strengthens existing connections in the direction of sensory layers to motor layers. This makes the network more feedforward and reflexive over time.

Modularity increases

Synaptogenesis also progressively increases the community structure of the brain as sub-networks for sensory or motor processing gradually emerge with maturation.

These principles have ontogenetic, phylogenetic, and functional implications discussed below.

The C. elegans wiring diagram is not stereotyped

We found that there was considerable variability in synaptic connectivity between this set of isogenic animals. About 43% of all connections and 16% of all synapses were not conserved between animals. This degree of variability contrasts with the widely held view that the C. elegans connectome is hardwired. In C. elegans, the idea that individual neurons should have identical connectivities probably stemmed from the fact that individual neurons are identifiable in each animal by virtue of their mostly stereotyped shape and lineage ^30,33. This stereotypy implies that their properties are genetically determined. Hence, if genetic regulation is strong, each identified neuron should have the same connections from one animal to the next. The original worm connectome, because it was assembled from partial datasets and only one complete adult nerve ring, could not address variability ²⁵.

We found that synaptic variability between animals was not uniform among cell types. For example, modulatory neurons have considerable variability in their output connections whereas motor neurons have little variability in their outputs. This contrast suggests that variability is in some way regulated between cell types and may be functionally important. For example, behavioral variability between animals can confer a fitness advantage to a population ³⁴. Synaptic variability may be a source of such behavioural variability, e.g., in the Drosophila visual system, variability among neurite morphologies has been linked to behavioural variability ³⁵.

In isogenic animals one mechanism for synaptic variability may be stochastic differences in gene expression. Variable expression levels have been observed even in the housekeeping genes in C. elegans embryos ³⁶. Neuronal activity can also be a driving force for synaptic remodeling. Individuals from an isogenic population reared in similar conditions will still experience differences in their local environments throughout life, a source of differences in neuronal activity that may translate into wiring variability ³⁷.

Developmental changes in the periphery of the connectome versus constancy in the central core

Why is interneuron connectivity so stable across maturation when compared to the sensory input and motor output of the brain? From an evolutionary standpoint it may not be surprising that the parts of an animal’s nervous system that interact with the outside world (mainly sensory systems and motor systems) are under high evolutionary pressure to maintain an animal’s fitness in changing environments. Such evolutionary changes in the nematode brain (phylogeny) may have accrued as developmental changes (ontogeny) in its wiring diagram.

The stability of the core parts of the nervous system across maturation implies that the central processing unit is robust enough to be used in different contexts. Maturation changes the flow of sensory information into the central processor and changes the readout of motor execution from the central processor without changing the central processor itself. Sensory maturation may reflect changes caused by learning and memory ³⁸. Motor circuit maturation may reflect adaptations to the changing musculature of the growing animal ³⁹.

The connectome becomes more feedforward during maturation

We observed an increased feedforward-bias of the adult brain that may be more effective in rapid information processing and making reflexive decisions. In contrast, the juvenile network with more feedback connections may have a greater capacity for learning and adaptation. Interestingly, feedback is what is used to train artificial neural networks that perform machine learning. After artificial networks achieve their desired performance, they operate in a feedforward manner. The architecture of the adult nematode brain may be a consequence of feedback-mediated optimization of it sensorimotor pathways.

The connectome becomes more modular during maturation

We observed an increased community structure of the brain’s network that suggests the emergence of specialized circuits with distinct roles. Each functional community emerges among neurons that are physically close to one another (Fig. 5g). In the nematode brain, the physical layout precedes the functional layout. Over time, neurons in proximity are more likely to acquire similar functionality by building their connectivity. The communities that are formed effectively create spatially compact areas for sensory or motor processing, reminiscent of distinct brain areas in larger animals.

Perspectives

In larger animals like insects and mammals that mature more slowly, synaptic remodeling can involve extensive changes in the nervous system. Apoptosis, neurite degeneration, and synaptic pruning can remove unwanted circuitry ⁴⁰. Cell proliferation and differentiation, neurite growth and guidance, and synapse formation can create new circuitry ⁴¹. In C. elegans, maturation must be fast and efficient. Each cell may be considered to be unique, and each is characterized by an intrinsic propensity for synaptic remodeling that occurs in the context of its stable morphology and fixed physical contacts with its neighbours. In light of these constraints, the nematode has evolved a broad set of principles for synaptic maturation to build its adult brain (Fig. 6).

In C. elegans, synaptic remodeling is extensive and widespread leading to changes from the cell to network level that likely has profound functional consequences on animal behaviour. Many investigations of flexibility in neural circuits and behaviour have focused on functional modulations of connectomes assumed to be anatomically static ^42,43. Our comparison of connectomes from birth to adulthood argues that the maturation of brain and behaviour cannot be separated from wiring changes. Comparative connectomics is needed to understand the origin of behavioural differences within and across species. High-throughput electron microscopy establishes a necessary foundation for understanding how genes, experience, and evolution create adult brain and behaviour.

Methods

Data acquisition

We studied wild-type (Bristol N2) animals reared in standard conditions: 35×10mm NGM-plates, fed by OP50 bacteria, and raised at 22.5 °C ⁴⁴. The animals were within a few generations of the original stock acquired from Caenorhabditis elegans Genetics Center (CGC) in 2001. All samples used in this study were derived from three batches of EM preparation.

Each EM sample was prepared and processed as previously described ²⁹ with small modifications to the substitution protocol of the last 3 datasets (in preparation). In short, isogenic samples reared in the same environment were high-pressure frozen (Leica HPM100 for datasets 1-5 and Leica ICE for datasets 6-8) at different stages of post-embryonic development. High-pressure freezing was followed by freeze-substitution in acetone containing 0.5% glutaraldehyde and 0.1% tannic acid, followed by 2% osmium tetroxide. For each life stage, we selected animals based on their overall size and morphology for EM analysis. The precise developmental age of each larval animal was determined based on its cellular compositions relative to its stereotyped cell lineage ³⁰, as well as the extent of neurite growth (see Supplemental Text). Three samples (datasets 2, 6, and 7) were prepared for transmission electron microscopy (TEM). Five samples (datasets 1, 3, 4, 5, and 8) were prepared for scanning electron microscopy (SEM).

For TEM, samples were manually sectioned at ∼50nm using a Leica UC7 ultramicrotome, collected on formvar-coated slot grids (Electron Microscopy Sciences, FF205-Cu), post-stained with 2% aqueous uranyl acetate and 0.1% Reynold’s lead citrate, and coated with a thin layer of carbon. Images were acquired using an FEI Techai 20 TEM and a Gatan Orius SC100 CCD camera.

For SEM, samples were serial sectioned at ∼30nm and collected using an automated tape-collecting ultramicrotome (ATUM) ⁴⁵. The tape was glued to silicon wafers, carbon coated, and sections post-stained with 0.5% uranyl acetate (Leica Ultrostain I, Leica Microsystems) and 3% lead citrate (Leica Ultrostain II, Leica Microsystems). Images were collected semiautomatically using custom software guiding a FEI Magellan XHR 400L ⁴⁶.

All images were acquired at 0.64-2 nm/px (∼25,000x). In total, these datasets comprise 94374 images, 5 teravoxels, and 2.4×10⁵ µm³. Images were aligned using TrakEM2⁴⁷ and imported into CATMAID ⁴⁸ for annotation.

All images will be made available on a public repository.

Connectome annotation

All cells within the brain were manually reconstructed by skeleton tracing in CATMAID ⁴⁸. The brain was defined as the nerve ring and ventral ganglion neuropil anterior of the ventral sub-lateral commissures. Chemical synapses were mapped manually. To reduce biases from different annotators, all datasets were annotated independently by three different people. Only synapses that were agreed to by at least two independent annotators were included in the final dataset.

Neurons were identified based on cell body position, neurite trajectory, and stereotypic morphological traits ²⁵. In the original connectome datasets as well as ours, some variability in cell body position was observed (see Supplemental text). However, every cell could be unambiguously identified in every dataset when all anatomical factors were taken into account. Negligible amounts of neuropil in our reconstructions could not be reliably identified as belonging to any known cell. These orphan fragments were relatively small (median length 0.38 µm) and rare (4.13±6.05 per dataset). Orphan fragments represent 0.18% of the total neurite length and 0.13% of all synapses, and were excluded from analysis.

Chemical synapses were identified by a characteristic presynaptic swelling containing a pool of clear vesicles adjacent to a dark active zone on the inside of the membrane ²⁹. Any cell adjacent to the active zone was identified as a postsynaptic partner. Presynaptic swellings were also typically characterized by mitochondria and cadherin-like junctions between pre- and post-synaptic cells ⁴⁹. A small fraction of postsynaptic partners exhibited postsynaptic densities.

Chemical synapses came in two varieties: classical synapses containing mostly clear synaptic vesicles surrounding the active zone and synapses of modulatory neurons containing mostly dense-core vesicles distant from the active zone. Most classical synapses also contained a small number of large dense-core vesicles at the periphery of the vesicle pool. Besides chemical synapses, neurons contained swellings with vesicles but no active zones. The majority of swellings of modulatory neurons did not have active zones ⁵⁰. These swellings were not annotated as synapses.

Final synapse annotations for all datasets are available at http://nemanode.org/.

Classification of neuron types

Neurons were classified as modulatory if they contained mostly large dark vesicles, or if they had been previously reported to use the neurotransmitters serotonin, dopamine, or octopamine ^51,52. Neurons were classified as motor neurons if they primarily made synapses onto muscles. Neurons were classified as sensory if they had specialized sensory processes and/or were previously reported to be have sensory capabilities. Neurons were classified as interneurons if most of their connections were to other neurons. Some neurons exhibit features corresponding to more than one type. These neurons were classified based on their most prominent feature (Table S1).

Volume segmentation for topological reconstruction

We computed the precise shape of every neurite in each EM image based on the skeleton tracing that was performed in CAT-MAID and a machine learning algorithm that recognized cellular boundaries. In brief, the algorithm expanded all skeleton nodes in each section until they fully filled the images of all labeled cells.

Cellular borders were predicted by a shallow Convolutional Neural Network (CNN) that builds on XNN ^53,54, a recently developed high performance system which computes convolutions on CPUs, to achieve border prediction throughput of ∼10MB/s ^55,56. Node expansion was computed with a dedicated Cilk-based code ⁵⁷ that parallelized the Dijkstra graph search algorithm. Code optimization allowed us to perform node expansion of an entire EM section in memory by a single multi-threaded process. Each software thread expanded an individual skeleton. Each pixel is attributed to a given cell by computing a generalized form of distance, taking into account the minimum number of cellular border pixels that must be traversed in a path connecting pixel and node. The generalized distance is computed using graph theory and concurrent data structures.

Volume traces were imported into VAST ⁵⁸ for manual proof-reading. At least 1,120 person-hours were spent proofreading the volumetric expansions.

Data processing for statistic analysis

Volumetric neuron traces were exported from VAST ⁵⁸ and imported into MATLAB. EM artefacts were manually corrected. To calculate the contact area of each cell pair, we performed two-dimensional morphological dilation of every traced segment across extracellular space until neighbouring segments made contact within 70 pixels (140-280nm). Expansion was restricted to the edge of the nerve ring. The total contact area was calculated as the sum of adjacent pixels for each segment in all sections. Contacts between cell bodies at the periphery of the neuropil were excluded.

Neuron skeletons and synapses were exported from CAT-MAID using custom Python scripts, and imported into Python or MATLAB environments for analyses. The module detection analysis was performed in MATLAB. Other analyses were implemented with custom Python scripts using SciPy and Statsmodels libraries for statistics. We excluded post-embryonically born neurons from our analyses.

For analyses related to neurites, both processes of neurons and muscles in the nerve ring were included. The neurite length was calculated using the smoothened skeleton of each neurite. The skeleton was smoothed by a moving average of 10 skele-ton nodes after correction of major alignment shifts. Spine-like protrusions were defined as any branch shorter than the 10% of the average neuron length. For analyses related to information flow, separating connections into feedforward, feedback, and recurrent, connections to muscles were excluded since they are all feedforward. All scripts and files used to generate all figures will be made available on a public repository.

Classification of connections

A total of 3113 connections (averaging 1292 per dataset) were assigned as stable, variable, or developmentally dynamic. 1647 weak connections (averaging 323 per dataset) had no more than two synapses in two or more datasets and were left-right asymmetric. These connections were classified as variable. The 1466 remaining connections were pooled by left-right cell pairs, resulting in 658 pair connections. The number of synapses in each pair connection was tested for relative increase or decrease across maturation (Spearman’s rank correlation, corrected for multiple comparisons using the Benjamini–Hochberg correction). Pair connections showing a significant change and at least a 5-fold change in synapse number from birth to adulthood were classified as developmentally dynamic. Remaining pair connections were considered stable if they were present in at least 7 datasets, and variable if present in fewer than 7 datasets.

Community structure analysis

Weighted stochastic blockmodeling (WSBM) ³² was used to define modules for individually for all eight connectomes. In this approach, modules are optimized on the likelihood of observing the actual network from the determined modules (log-likelihood score) based on two exponential family distributions. We chose the probability of establishing connections to follow a Bernoulli distribution and the synapse number for each connection to follow an exponential distribution. These distributions fit the data best according to the log-likelihood score and resulted in left-right cell pairs being assigned to the same modules.

In order to find a stable and representative number of modules for each connectome, we used a consensus-based model-fitting approach, similar to previously described ⁵⁹. First, to ensure unbiased coverage of the parameter space, we fitted the model independently 300 times using an uninformative prior for each potential number of modules (k = 1, …, 10). This procedure was repeated 100 times to yield a collection of models with concentrated and unimodally distributed log evidence scores. To improve the stability of the models on multiple runs, we increased the parameters for a maximum number of internal iterations to 100. For each dataset, we chose the number of modules whose collection of models had the highest mean posterior log-likelihood score. If scores for two different numbers of modules had significant overlap, the number of modules closest to the connectome at an earlier developmental timepoint was selected.

Finally, for each dataset we found a representative consensus module assignment that summarized all 100 models ⁵⁹. In brief, considering all 100 models, we calculated the frequency of each cell being assigned to each module, and used this as a new prior to fit another 100 models. This procedure was repeated until convergence, when the consistency of the 100 models was larger than 0.95.

Community structure validation

We validated the community structure defined by WSBM using a previously described method ⁵⁹. In brief, for each possible number of modules k = 1, …, 10, the quality of the best final model determined by WSBM was examined to validate the model chosen by the log-likelihood score. For each k, we fit a WSBM model with a prior matching the module assignment, and reverse simulated 2000 synthetic connectomes from the model. For each synthetic connectome, we recorded 8 statistic measurements: degree distribution, in-degree distribution, out-degree distribution, weight distribution, in-weight distribution, out-weight distribution, betweenness centrality, and weighted clustering coefficient. These distributions were compared to the actual connectome using a Kolmogorov–Smirnov (KS) statistic test, and summarized by computing the mean KS energy, defined as the mean value of all 8 KS statistic values. A lower mean KS energy indicated a better match. For the connectomes of early developmental stages, an equal match was found for k = 3 … 6 (Fig. S8b). For the adult connectomes, k = 6 matched the connectome significantly better than k < 6 (Fig. S8b).

AUTHOR CONTRIBUTIONS

J.W.L, A.D.T.S., and M.Z. conceived the study. Y.M., R.P., and N.S. designed the algorithm for automated volumetric reconstruction ( for correspondence). D.R.B. designed the pipeline for automated EM acquisition ( for correspondence). Y.W. designed software for EM alignment ( for correspondence). D.W., B.M., J.K.M., D.H., R.L.S, and M.Z. generated and imaged most of the electron micrographs. D.W., B.M., and J.K.M. performed most annotation. D.W. performed most analysis. D.R.B., W.X.K., and Y.L. performed additional experiments and analysis. A.D.C. guided early cell identification and annotation. D.W., J.W.L, A.D.T.S., and M.Z. wrote the manuscript. All authors discussed the results and reviewed the manuscript.

COMPETING INTERESTS

The authors declare no competing interests.

Supplemental figures

Table S1. Cell types in the nerve ring.

Table S2. Members of communities detected by WSBM colored by type.

Table S3. Optimal number of communities detected by WSBM using subsets of connections.

Figure S1. Topological models for seven C. elegans brains at respective developmental stages.

All models include the complete neuropil of the brain, consisting of the nerve ring and ventral ganglion. Cells are colored by type.

Figure S2. Closeup of an adult brain connectome.

Wiring diagrams for an adult connectome (dataset 8). Each circle represents a cell. Circle color denotes cell type. Each line represents a connection with at least one chemical synapse between two cells. Line width indicates synapse number. Straight lines direct information from sensory to muscle layers whereas curved lines direct information in reverse. Cell coordinates are represented as in Fig. 1b, with overlapping cells manually separated.

Figure S3. Neurites grow while maintaining brain topology.

a. 3D reconstructions of one neuron class (AIZL and AIZR) across maturation. The overall topology was maintained, whereas the number of spine-like protrusions (grey arrows) increased over time. b. Correlation of the relative neurite length of each branch between L1 (dataset 1) and adult (dataset 8). The length of each neurite is normalized against the total neurite length of the neuron. c. Proportion of postsynaptic spine-like protrusions increases with maturation. *** p < 0.001, Spearman’s rank correlation. d. Total number of spine-like protrusions in the brain increases almost 5-fold with maturation. *** p < 0.001, Spearman’s rank correlation. e. Proportion of synapses that have at least one spine-like protrusion postsynaptically increases with maturation. *** p < 0.001, Spearman’s rank correlation. f. Connectivity asymmetry decreases from birth to adulthood, most significantly during L1. Asymmetry is defined as the coefficient of variation (CV) in synapse number between left-right cell pairs. g. Total number of missing connections decreases from birth to adulthood, most significantly during L1. A missing connection is defined as a connection absent in only one dataset and from one side of the brain.

Figure S4. Three neuron classes grow new neurites after birth.

Topological models of ADE, SAAV, SAAD, and RIM in L1 (dataset 2), L3 (dataset 6), and adult (dataset 8). These neurons pairs grow new major branches, highlighted by dotted gray circles. The new branches of ADE sprout outside the nerve ring.

Figure S5. Non-uniform distribution of connections and strengthening of connections across maturation

a. Distribution of the total number of input and output connections per neuron at birth. Some neurons have more connections than others.b. The relative number of synapses added to existing connections is correlated between outputs of the same cell compared to connections to and from different cells. The relative number of synapses added is quantified as the fold increase of synapse number from birth (dataset 1) to adulthood (averaged between datasets 7 and 8). ** p < 0.01, *** p < 0.001, Mann–Whitney U test, FDR adjusted using Benjamini–Hochberg correction.

Figure S6. Connectivity matrix of the C. elegans brain throughout maturation.

A connectivity matrix that includes all connections observed in eight C. elegans brains. Cells are pooled by left-right pairs. The size of each connection represents its largest synapse number in any dataset. Stable, developmentally dynamic, and variable connections are colour-coded by their classification (see Methods).

Figure S7. Propensity of forming variable connections correlates with cell type.

a. No synapse number provides a good filter for specific removal of variable connections. Any arbitrary threshold removes both variable and stable connections. Total number of both variable connections and non-variable (stable and developmentally dynamic) connections upon filtering by different synapse numbers. b. Arbitrary thresholding connections by synapse number leaves substantial proportion of variable connections for all cell types. The non-uniform distribution of variable connections is consistent when weak connections are excluded. c. The low variability of connections from motor neurons to muscles cannot be simply explained by saturation of their physical contacts by synapses. Physical contacts are not saturated for connections for any cell type. Motor neurons, which have the lowest proportion of variable connections (Fig. 4b), are not restricted by few available potential synaptic partners. d-g. Higher variability for certain cell types could also not be simply explained by a fixed probability of an erroneous connection by neurons that exhibit abundant synapse formation. d. The number of variable connections formed by a cell does not correlate with the strength of its stable output connections. Each data point represents one cell. ns, not significant, Spearman’s rank correlation coefficient. e. The number of synapses for stable output connections by cell types. Modulatory neurons, which exhibit a higher proportion of variable connections than other cell types (Fig. 4b), do not exhibit more synapses per stable output connection than other cell types. f. The number of variable connections formed by a cell does not correlate with the number of synapses added to existing stable output connections from birth to adulthood. The relative number of synapses added is quantified as the fold increase of synapse number from birth (dataset 1) to adulthood (averaged between datasets 7 and 8). Each data point represents one cell. ns, not significant, Spearman’s rank correlation coefficient. g. The relative number of synapses added to existing stable output connections by cell types. Connections from modulatory neurons, which have the higher proportion of variable connections that other cell types (Fig. 4b), do not exhibit higher increase in synapse number than other cell types. For all panels, the synapse number for the adult brain (averaged between datasets 7 and 8) is shown.

Figure S8. Cell communities across maturation.

a. The log-likelihood score for each WSBM model (see Methods). b. The deviation between the connectome and each synthetic network generated from the best WSBM model, measured by the mean KS energy (see Methods). A lower deviation indicates a better match between the actual connectome and network generated from the model. Adult datasets show a clear preference to more than 5 modules, while juvenile datasets do not.

Supplemental text

Estimation of the developmental age of datasets

The developmental age of each sample was established based on the described temporal cell division pattern exhibited by wild-type (N2) larva raised at 25°C ³⁰. Dataset 1: L1 at birth. No Q cell division, which occurs ∼3 hours post-hatching (hrs) after birth, and Q cell nuclei are symmetrical, before nuclei migration, which occurs ∼2 hrs after birth, placing this sample so close to birth, at ∼0 hrs. Dataset 2: L1 at 5 hrs. The lack of H1 division, which occurs at∼7.5 hrs after birth, and no growth of PVC and SAA posterior neurites placed this sample to be ∼5 hrs after birth. Dataset 3: L1 at 8 hrs. With H1 just completing its division and P5/6 starting their migration, this sample was placed at ∼8 hrs after birth, when both events take place. Dataset 4: L1 at the very end of the larvel stage (16hr). This sample was estimated to be 16hrs, near the end of the L1 lethargus. It has two layers of cuticle. Both P11.aaa and P12.aaa have divided. V5R.p is in the midst of division, and H1.a has not yet divided; all happen at ∼16 hrs. Dataset 5: L2 towards the end of the larval stage (23hr). SML/R have not divided, which occurs at ∼29 hrs. It has 40 gonad cells, and a slight double cuticle that indicates the end of L2. However, its gut lumen contains food, placing this sample shortly before entering L2 lethargus, which occurs at ∼16 hrs. Dataset 6: L3 at 27 hrs. Based on the partial outgrowth of the RMF neurites, which is born at 23 hrs, this sample was estimated to be ∼27 hrs. Datasets 7 and 8: Young adults at 45hr. Both samples have adult cuticles but are relatively small compared to other adults. The exact ages of the two young adult samples are uncertain, so they are treated as equals for analyses.

Anatomical inconsistencies between samples

A few major anatomic inconsistencies are observed in some datasets, likely due to heterogeneity or imprecision of development processes. These events do not have an impact on overall connectivity, as all non-variable connections between individual neuron classes were conserved. Dataset 2: CEPDL cell body is shifted to the anterior ganglion. Dataset 3: RIFL neurite terminates prematurely laterally, not reaching the dorsal midline. Dataset 4: RIH cell body is shifted to the anterior ganglion. PVCL and PVCR neurites both go right-handedly around the nerve ring, appearing as PVCR. Dataset 5: RMHL and RMHR neurites both transverse right-handedly around the nerve ring, appearing as RMHR. ADAL terminates prematurely at a dorsal sub-lateral,position, not reaching the dorsal midline. Dataset 6: PVR neurite is fragmented. Dataset 7: RIFL and RIFR neurites both transverse right-handedly around the nerve ring, appearing as RIFR.

ACKNOWLEDGEMENTS

We thank Valeriya Laskova for assistance in developing the EM sample preparation protocol. We thank Bob Harris for assistance with high-pressure freezing. We thank Marianna Neubauer, David Kersen, Anabelle Paulino, Manusnan Suriyalaksh, Amelia Srajer, Maggie Chang, Sean Ihn for help with imaging. We thank Ignacio Arganda-Carreras and Jenny Qian for guidance on with EM alignment. We thank Steven Cook, Christine Rehaluk, and Mona Wang for synapse annotation in some datasets. We thank Jade Ho, Christopher Morii-Sciolla, Isis So, Min Wu and Chi Yip Ho for help with generating ground truth and proofreading for topological reconstruction. We thank Alexander Mateev, Lu Mi, and Hayk Saribekyan for help generating and applying algorithms used in this project. We thank Jerry Wang and Danqian Cao for help with statistical analyses. We thank Albert Lin, Chris Tabone, and Vivek Venkatachalam for setting up and supporting the server for synapse annotation. We thank Soomin Maeng and Dylan Fong for assistance with the develop of www.nemanode.org. We thank members of the Zhen, Samuel, and Lichtman labs for comments. We especially thank Guangwei Si for critical reading and suggestions. We thank David Hall, Jagan Srinivasan, and Albert Cardona for early advice in this project.

J.K.M. was supported by National Science Foundation Physics of Living Systems (NSF 1806818). B.M. was supported by the Mount Sinai Foundation. J.W.L., A.D.T.S., and M.Z. were supported by the Human Frontier Science Program (RGP0051/2014). A.D.T.S and M.Z. were supported by the National Institutes of Health (R01-NS082525-01A1). A.D.T.S. was supported by National Institutes of Health Brain Initiative (1U01NS111697-01) and National Science Foundation BRAIN EAGER (IOS-1452593). M.Z. was supported by Canadian Institutes of Health Research (MOP-123250 and Foundation Scheme 154274), the Radcliffe Institute, and the Mount Sinai Foundation.

Footnotes

• http://nemanode.org/