Introduction

Neural circuit structure supports function. The underlying image data that yields anatomical connectomes (or wiring diagrams) are typically obtained using volume electron microscopy (vEM) techniques (Collinson et al., 2023). Since the first complete connectome was published for C. elegans (White et al., 1986), these last decades have seen an increase in the generation of vEM datasets, as reviewed in (Kaiser, 2023) and others. The expansion in available anatomical connectomes has resulted from recent advancements in: 1) data generation (via automation of EM data acquisition (Xu et al., 2017; Eberle and Zeidler, 2018; Zheng et al., 2018; Phelps et al., 2021); and 2) alignment, segmentation and reconstruction (including recent implementation of AI-driven methods) as reviewed in (Galili et al., 2022; Choi et al., 2024) and others. As these developing methodologies continue to improve, they will continue to facilitate the generation of additional connectomes of whole brains and organisms.

The increasing availability of vEM datasets, including the first series of developmental connectomes published for C. elegans (Witvliet et al., 2021) has highlighted the need for new tools to enable intuitive examination and comparisons across connectomes to promote novel discoveries (Kasthuri et al., 2015; Lichtman et al., 2014; Barabási et al., 2023; Xu et al., 2021). It has also underscored the fact that vEM datasets contain a wealth of untapped information that has yet to be fully examined, represented and integrated for more comprehensive analyses (Perez et al., 2014; Brittin et al., 2021). For example, vEM datasets enable nanoscale explorations of the underlying cell biological features that govern the properties of neural circuit architectures (Rivlin et al., 2024; Brittin et al., 2021; Moyle et al., 2021; Witvliet et al., 2021; Heinrich et al., 2021). Yet most of these cell biological features (cell morphologies, contact profiles, organelle positions and shapes, etc) are not currently represented in most anatomical connectomes. Quantification of cell biological data result in high-dimensional datasets that require new approaches for their analyses and representations. The advances in vEM data generation and the resulting need for new methodologies in data science and integrated representations of neuronal relationships (e.g. from neuronal positions to neuropil structures) is akin to how advances in genetic sequencing required new methodologies in bioinformatics and new, integrated representations of genomic data (e.g. from gene sequence to gene structure) (Swanson and Lichtman, 2016). Addressing this gap holds the promise of integrating new knowledge from the fields of cell biology, neurodevelopment, physiology and systems neuroscience towards explaining how nervous system structure underpins its function.

Most representations of anatomical connectomes have focused on defining neuronal relationships at the level of the chemical synapse (NemaNode; WormWiring; EleganSign; FlyWire) (Witvliet et al., 2021; Cook et al., 2019; Fenyves et al., 2020; Dorkenwald et al., 2023). While the existence of chemical synapses between neuron pairs is an important feature of neuronal communication, these representations do not capture other neuroanatomical features that also underlie neuron structure and function, including contact sites from adjacent (or nearby) neurons. Recent work in C. elegans examined neuronal relationships by quantifying neuron-neuron contact sites to build contact profiles, or contactomes (Brittin et al., 2021). Examination of the contactome with data science approaches uncovered structural principles that were not evident from interrogating the synaptic connectome alone (Moyle et al., 2021; Brittin et al., 2021). These included the existence of higher- order structural motifs and the stratification of neurons (Moyle et al., 2021), whose hierarchical assembly during development is guided by centrally located pioneer neurons (Rapti et al., 2017). Moreover, integrating neuronal adjacencies (contactome) with synaptic profiles (connectome) allowed for a deeper understanding of the functional segregation of neurons within the stratified neuropil structures (Brittin et al., 2021; Moyle et al., 2021). Key to achieving this were data science approaches such as Diffusion Condensation (DC) and C-PHATE (Brugnone et al., 2019; Moon et al., 2019), which resulted in reduced dimensionality of the neuronal relationships, revealing architectural motifs across various scales of granularity, from individual neurons within circuits, to individual circuits within the neuropil. These techniques produced graphs that enabled exploration of these computationally identified groups (Moyle et al., 2021). DC/C-PHATE graphs are powerful tools, but they have yet to be integrated to connectomics datasets as to enable explorations of the underlying cell biological features. This limits their effectiveness for hypothesis generation and comparative analyses across connectomes.

To address this, we generated NeuroSCAN, a tool for exploring neuroarchitectures across vEM datasets via novel representations of the connectome, contactome, and anatomical networks. Neu-roSCAN is an online, open-source platform that facilitates comparisons of neuronal features and relationships across vEM data to catalyze new insights of the relationships that underpin architectural and functional motifs of the nerve ring neuropil. NeuroSCAN builds on recent publications in whole-brain EM datasets, integrating the latest set of developmental connectomes (Witvliet et al., 2021) and employing data science tools (Brugnone et al., 2019; Moon et al., 2019) to examine neuronal relationships based on contact profiles. We demonstrate how these integrated representations of neuronal relationships facilitate comparisons across these connectomes, catalyzing new insights on their structure-function and changes during development. NeuroSCAN achieves this by addressing three challenges in current neuronal representations: 1) accessibility of specific neuronal cell biological features (i. e. synapses and contacts), 2) integration of features for examining neuronal relationships across anatomical scales, and 3) spatiotemporal comparisons of these features across developmental datasets.

These challenges were addressed by 1) creating representations of contact sites and establishing the ability to visualize subsets of synaptic sites; 2) enabling synchronous visualization of neuron morphologies, contacts and synapses and integrating these cell biological features with algorithmically-generated graphical representations of neuronal relationships; and 3) enabling simultaneous exploration of these relational representations across developmental connectomes. NeuroSCAN was designed as a suite of tools that facilitates future incorporation of additional datasets and representations with the goal of enabling integrated data exploration beyond the available C. elegans connectomes. The NeuroSCAN-based approaches used here for C. elegans could be applicable to other systems as new EM-based datasets and reconstructions become available.

Results (Comparing contactome-based relationships using C-PHATE.)

The adult hermaphrodite C. elegans nerve ring is a neuropil of 181 neurons of known identities, morphologies, contact profiles, and synaptic partners (White et al., 1986). Even for this relatively small neuropil, representations of a single feature type, such as neuronal contact profiles, constitute over 100,000 data points of multidimensional information: cell identity, region of contact, presence of synapses, etc. Analysis of this multidimensional information requires approaches that can both capture higher order patterns of organization while enabling researchers to access the underlying cell biological features resulting in these relationships. We implemented Diffusion Condensation (DC), a clustering algorithm that iteratively group neurons based on the quantitative similarities of their ‘contact’ or ‘adjacency’ profiles (Brugnone et al., 2019; Moyle et al., 2021). Briefly, DC makes use of pair-wise quantifications of adjacent neuron contacts to move neurons with similar adjacency profiles closer together by applying a diffusion filter in a multidimensional manifold. This diffusion filter effectively removes variability between neurons at each iteration. As iterations proceed, individual neurons (and eventually groups of neurons) are clustered together based on how close their diffused contact profiles are to one another in the manifold (Brugnone et al., 2019). In this way, DC uncovers hierarchical neuronal relationships in the contactome (Moyle et al., 2021).

To ensure accurate comparisons of DC across available EM datasets (Witvliet et al., 2021; White et al., 1986), we first empirically set minimum-distance adjacency thresholds (measured in pixels; Supplemental Table 1) to build adjacency profiles (see also Methods and Materials), schematized in Figure 1A-C). We then quantified the lengths of the physical adjacencies (or contacts) between neuron pairs (Brittin et al., 2021) and built an adjacency matrix for each of the seven selected C. elegans contactome datasets (L1, 0 hours post hatch (hph); L1, 5hph; L2, 16hph; L3, 27hph; L4, 36hph; Adult 48hph (Figure 1C; See also Methods and Materials). To visualize and compare the results from DC, we used C-PHATE (Moon et al., 2019; Moyle et al., 2021), a 3-D visualization tool that builds a hierarchical visual representation of the DC agglomeration procedure (Figure 1D-E). In C-PHATE visualizations, the DC output is mapped in 3-D space with spheres. Initially, all individual neurons in the neuropil dataset are at the periphery of the C-PHATE graph (left hand side in schematic in Figure 1D, edges of graph in Figure 1E). Neurons are iteratively condensed based on the similarity of their adjacency profiles (schematized in Figure 1D). In the last iteration of DC, there is a single point at the center of the C-PHATE graph which contains the entire neuropil (Figure 1E, red dot). C-PHATE representations enable visualization and comparisons of contactomes across datasets, and explorations of neuronal relationships, from individual neuron interactions to circuitcircuit bundling (Figure 1F and Figure 2).

Figure 1.

Figure 2.

By Larval stage 1 (L1), neuronal differentiation has concluded and 90% of the neurons in the neuropil (161 neurons out of the 181 neurons) have entered the nerve ring and adopted characteristic morphologies and positions (Sun and Hobert, 2023). Although the organism grows approximately 5 fold from L1 to the adult, contacts in the nerve ring are also largely established by L1 and preserved during postembryonic growth (Witvliet et al., 2021). In agreement with this, when we used DC and C-PHATE to examine contactomes from these datasets we consistently identified four main superclusters– Stratum 1, Stratum 2, Stratum 3, and Stratum 4 (Figure 2B-F). These findings are consistent with previous studies on the Larval Stage 4 (L4) and adult contactomes (Moyle et al., 2021), and further suggest that neurons establish core relationships during embryogenesis and maintain them into adulthood. Moreover, aligning the neuronal morphologies of strata members reveals a persistent layered organization to the nerve ring neuropil (Figure 2 G-K), and exploring the functional identities of the neurons in each stratum suggests that there is spatial segregation of sensory information and motor outputs (see (Moyle et al., 2021) see: Supplemental Tables 3, 4, 5, 6. Our findings are in agreement with previous reports that the organization of the nerve ring is largely established in embryogenesis, and then maintained during postembryonic growth (Witvliet et al., 2021). Our findings also demonstrate the utility of DC and C-PHATE analyses in extracting, visualizing and comparing the structure of the neuropil architecture across contactomes.

Because DC and C-PHATE allow for the examination of relationships at varying levels of granularity, they also facilitate the interrogation of the architectural motifs that underlie distinct neural strata. A more detailed examination of clusters reveals that while the overall strata are preserved, the underlying neuronal configurations undergo changes during post embryonic growth (Figure 2 B-F, Figure 3, see: Supplemental Tables 3, 4, 5, 6). Three general features were extracted from these analyses: 1) individual neurons renegotiate their positions in the context of the identified C-PHATE clusters in different developmental contactomes, suggesting developmental changes; 2) the degree of these changes varied across the distinct strata; and 3) the degree of these changes mapped onto known features of each stratum, such as plasticity. For example, Stratum 1, which contains shallow reflex circuits, displayed the fewest changes among the developmental connectomes (Figure 3 B-F; Supplemental Table 3). On the other hand, Strata containing circuits associated with behavioral plasticity (Stratum 3 and Stratum 4), displayed the largest changes across postembryonic development (Figure 3H-L; Supplemental Tables 5, 6).

Figure 3.

To examine the changes in DC/C-PHATE during postembryonic development, we made the C-PHATE plots fully interactive. This enables users to hover over and identify members of each intermediate cluster, to highlight specific cell trajectories via pseudo-coloring, and compare specific neuronal relationship dynamics across development within a multiview window of distinct C-PHATE plots (Figure 1 E-F, Supplemental Figure 6, Supplemental Video 1). Because C-PHATE graphs ultimately represent cells of known identities, we reasoned that interactive mapping of the C-PHATE cluster objects to their component cellular identities and anatomies could yield greater insights on neurodevelopmental changes, linking the algorithmic abstractions of the relationships with the cell biological features and their changes across development (Figure 4).

Figure 4.

To examine our hypothesis and determine the utility of C-PHATE for discovery, we inspected specific regions where the distribution, or ‘shape’ of superclusters changed across the set of developmental contactomes. When comparing C-PHATE graphs representing distinct contactomes, we accounted for changes in the iterations at which “merge events” (or co-clustering of neurons) occurred. The logic in considering the ‘iterations of the merge events’ is because variations in contact profiles influence changes in iterations of merge events. Based on these criteria, we focused on a region displaying changes in Strata 3 and 4 and using the interactive C-PHATE graphs (Figure 4 A-E), we determined the identities of neurons that changed clustering patterns across the developmental contactomes (Figure 4F and G). Specifically, we focused on two interneurons, named AIML and PVQL, which we observed undergo a change in their cluster assignment from Stratum 4 (at L1) to Stratum 3 (at Larva stage 4, L4; Figure 4A and D). We pseudo-colored the trajectories of the AIML and PVQL neurons in C-PHATE to explore the changes in their merge events throughout the developmental stages (Figure 4F and G, Supplemental Figure 1, Supplementary Table 7). Comparing L1, L2 (Larval Stage 2) and L3 (Larval Stage 3) datasets, we observe the AIML and PVQL neurons merge at iterations 16, 14 and 22 (respectively). The increasing numbers of iterations across the L1, L2 and L3 datasets suggests the relative contact profiles of AIML and PVQL diverge across these contactomes (Figure 4F and G; Supplemental Figure 1; Supplementary Table 7). Yet, between the L4 and Adult datasets, we observe the PVQL and the AIML neurons merge at iterations 20 in the L4 and iteration 14 in the Adult (Supplemental Figure 1; Supplementary Table 7). The decrease in the number of iterations required for the merge event suggests that the relative contact relationships of AIML and PVQL eventually converge between L4 and adult animals. Comparison of the identities of the neurons that co-cluster with AIML and PVQL similarly suggests that the contact relationships varied across developmental stages (Figure 4F and G, Supplemental Figure 1, Supplementary Table 7).

Visualizing contact profiles in individual cells

DC/C-PHATE changes should result from changes in contact profiles. To link the observed changes in the C-PHATE graphs with the cell-biological changes in contact profiles, we generated a tool that would simultaneously enable: 1) 3D visualization of the cell-cell contact sites onto individual neuronal morphologies; 2) examination and comparisons of these contact profiles throughout development for the available contactomes; and 3) integration with DC/C-PHATE to link C-PHATE cluster objects to the 3-D morphologies of the algorithmically clustered cells. With these capabilities integrated, we could simultaneously view the contactome from two complementary perspectives – at an abstract systems level via DC/C-PHATE and at a cell biological level via 3D contact modeling – to perceive the architectural themes that underlie similar network patterns.

To create this tool, we generated 3D models of the area of physical contact between adjacent neuron pairs (Supplemental Tables 1, 2, Methods and Materials; Figure 5) Supplemental Figure 2). Visualizing contacts from all adjacent neurons builds a multi-colored skeleton of the neuron morphology mapped onto the boundaries of this neuron (Figure 5A and C). Because the identities of the neurons are known and linked to the 3D contact models, we built text pop ups that define the contact partners for each site (Figure 5C). Furthermore, since neuron names are consistent across EM datasets, we can link and compare contact sites across development (Figure 5D). Additionally, we can analyze the representations of contact sites in the context of DC/C-PHATE clustering profiles (Figure 5B), 3D models of neuronal morphologies (Figure 1F), and 3D models of synaptic sites for any neuron(s) across development (Figure 7).

Figure 5.

We used the integrated tools of DC/C-PHATE and 3D representations of the contact profiles to examine the potential cell biological changes leading to the DC/C-PHATE clustering changes observed for the AIML neuron during development. With these tools, we observed changes in the identities of the contacts made in the dorsal region of the AIML neurite (Figure 5D; Supplemental Figure 3). Specifically, in the L2 stage (as compared to L1), we observed a decrease in the contacts from PVQL and an increase in contacts from the AVF neurons. This change persists to the adult stage (Figure 5D; Supplemental Figure 3).

To then determine the possible source of these developmental changes in contacts, we visualized 3D models of the segmented morphologies for these neurons across L1 to adulthood (Figure 6). We find that AIM and PVQ neurons maintain similar morphologies throughout development (Figure 6C), while AVF neurons undergo substantial neurite outgrowth onto new regions of contact between AIM and PVQ (Figure 6 B-D). Specifically, the data revealed that although the AVF neurons terminally differentiate in the embryo, they do not grow into the nerve ring until the L2 stage, and continue to grow until the Adult stage (Figure 6 B-D). The AVF neurons grow in between the AIM and PVQ neurons (Figure 6D), altering their contact profiles, which likely contributes to the observed changes in the C-PHATE graphs (although we note that DC/C-PHATE representations systematically cluster neurons based on relative similarities across contact profiles, not solely by scoring changes in specific contacts within any given pair (Figure 4F and G; Figure 5B and D; Supplemental Video 2). We also observe that both AVFL and AVFR grow into the nerve ring alongside AIML, later continuing to grow around to reach AIMR, and that these relationships were also reflected in the C-PHATE graphs in terms of the clustering profiles throughout development; (Figure 4G; Supplemental Figure 1).

Figure 6.

We then examined if the developmental changes in contact profiles result in changes in circuitry. We examined this by layering on synaptic information. Despite dwindling AIM-PVQ contacts, AIM and PVQ neurons maintained their synaptic relationship throughout development, with synaptic sites observed primarily at the base of AIM neurons, a region of persistent contact with PVQ (Figure 7A-B). We observed that increases in contacts between AIM and AVF neurons resulted in additional en passant synapses at the new points of contact, beginning at the L2 stage and continuing to adulthood (Figure 7A-B). We also observed that AVF forms synapses with the adjacent PVQ neurons (Figure 7; Supplemental Figure 4).

Figure 7.

In summary, by integrating, representing and comparing datasets using the new C-PHATE tools and contact profiles in NeuroSCAN, we identified developmental changes in the relationships of AIM, AVF and PVQ. This case-study highlights the utility of combining cell biological representations (such as morphologies, contacts and synapses) with coarse-grained systems-level representations (like DC/C-PHATE) of vEM datasets to uncover developmental changes which could be further explored experimentally. Therefore, NeuroSCAN serves as a powerful platform for generating hypotheses for empirical testing, which can lead to insights into the dynamics of circuit development.

NeuroSCAN: Facilitating multi-layered interrogation of neuronal relationships in the C. elegans nerve ring throughout larval development

NeuroSCAN is built as a web-based client-server system designed to enable the sharing of anatomical connectomics data with an emphasis on facilitating the analyses of neuropil relationships across hierarchies and scales. To achieve this, we integrated tools of neuroanatomical investigation from the available C. elegans nerve ring connectomes and contactomes with a collection of 3-D modeled elements (morphologies, contacts and synapses and C-PHATE) representing different aspects of neuronal architecture and relationships (Figure 8). NeuroSCAN differs from other available web-based tools in this area with the integration of C-PHATE graphs that enable exploration of hierarchical organizations of stratified fascicles, the availability of new tools to examine the contactome, and the integration of these data with existing connectome and morphological datasets across developmental stages.

Figure 8.

NeuroSCAN has eight key user-driven features: (1) C-PHATE, with the ability to highlight clusters containing neurons of interest (Supplemental Figure 6, Supplemental Video 1), (2) reconstructions of neuronal morphologies (Supplemental Figure 10, Supplemental Video 3) (3) reconstructions of neuronal morphologies of C-PHATE cluster members with a right-click on C-PHATE clusters (Sup-plemental Video 1), (4) 3-D renderings of neuronal contacts to visualize the spatial distribution of contact profiles (Supplemental Figure 5, Supplemental Video 4) (5) 3-D representations of synaptic sites with the option to visualize subsets of those sites (Supplemental Figure 7, Supplemental Video 4) (6) the ability to perform side-by-side comparisons across development (Supplemental Figure 11, Supplemental Video 3), (7) the option to pseudo color each object to highlight points of interest (Supplemental Figure 11, Supplemental Video 3) and (8) each item is an individual object with the ability to be further customized by the user (Supplemental Figures 11, 12).

The NeuroSCAN website architecture and data structure were designed to integrate these key user-driven features via a modular platform and linked datasets. The architecture uses Geppetto, an open-source platform designed for neuroscience applications, modularity, and large datasets (Cantarelli et al., 2018). Briefly, the architecture is effectively separated into two applications, a frontend React/JavaScript bundle that is delivered to the client, rendering the neuron data and assets, and a NodeJS application that exposes a JSON API, serving the neuron data and assets based on user interactions (Supplemental Figure 13). The backend uses a Postgres Database to store underlying data (Supplemental Figure 14), a Persistent Storage Volume that houses and serves static assets, and a variable number of Virtual Machines to run the frontend and backend application code, scaling as needed to accommodate traffic. The User Interface is a React application that allows users to filter, sort, and search through the Neurons so that they can be added to an interactive canvas (Supplemental Figure 13). When users add Neurons to a viewer, a .gltf file is loaded in for a given model (Synapses, Neurons, Contacts) at the selected developmental stage (Supplemental Figure 13), which can then be manipulated in the 3D environment or layered with other meshes as needed. NeuroSCAN can be used on common web-browsers (e.g. Google Chrome, Safari) and mobile devices.

The underlying data model makes use of tables representing Synapses, Neurons, Contacts and Developmental Stages. Relationships between these models are represented by foreign keys (Supplemental Figure 14). Source data is defined in a file-tree structure containing various assets (such as .gltf files representing various entities), as well as CSV’s which store relationships across entities. The directory structure outlines a vertical hierarchy, starting at the developmental stages, then branching downwards onto neuron and synapse data. A Python script is invoked to traverse the directory tree and parse the files, writing to the database accordingly. This configuration enables: 1) verification of the ingested data and 2) quick search times through the datasets to identify related items. Code is version-controlled in GitHub (https://github.com/colonramoslab/NeuroSCAN) and deployed through a CI/CD pipeline when updates are committed to the main branch (Supplemental Figure 13).

NeuroSCAN: practical considerations

We offer seven practical considerations for users. First, NeuroSCAN is available on mobile platforms as a quick and convenient way to look up neuron morphologies and relationships. Second, since contact sites offer the ability to explore the surrounding neurons and the position(s) of contact between adjacent neurons, NeuroSCAN is designed to enable studies of adjacent neurons (e.g. phenotypes that result in site-specific ectopic synapses; neuron morphology changes that may affect specific surrounding neurons; developmental events requiring communication between neurons, etc.). Third, C-PHATE can be used to identify neurons with similar contact profiles. Because contact profiles are associated with circuits identities (Moyle et al., 2021), exploration of neuronal relationships via C-PHATE can be used to identify new relationships between specific neurons and circuits. Fourth, visualization of subsets of synaptic and contact sites allows direct comparisons to light microscopy approaches such as cell-specific labeling of synapses or GFP-Reconstitution across synaptic partners (Feinberg et al. 2008). Fifth, because the color and transparency of each 3-D model can be customized, users can further integrate NeuroSCAN outputs of additional atlases (for gene expression, neurotransmitter and receptor expression, functional connectivity, etc. (Packer et al., 2019; Taylor et al., 2021; Wang et al., 2023; Fenyves et al., 2020; Randi et al., 2023) and directly use the NeuroSCAN outputs to create figures and comparisons (as done for this paper). Sixth, although synaptic sites with BWM (body wall muscles) are included in NeuroSCAN, the current data model limits the ability to search for these non-neuronal cells. Users can search for neurons with synapses to BWM to find this datatype. Seventh, to enable direct comparisons between our data representations and the primary EM data, the original annotations have been preserved and can be accessed by users via the sister app, CytoSHOW (CytoSHOW.org). As the data continues to be curated, the modular design of NeuroSCAN and its companionship with CytoSHOW enables integration of future annotations.

Discussion

NeuroSCAN is an integrative tool for analyzing detailed, web-based representations of neuronal connectomes and contactomes throughout post-embryonic development in C. elegans. Connectomes and contactomes are derived from volume electron microscopy (vEM) micrographs of neuropil regions (Witvliet et al., 2021; White et al., 1986). These EM micrographs are information-rich and have the potential to reveal architectural motifs across scales, from the nanoarchitecture of the neuron to the neuroanatomy of each circuit in the brain. Cell biological features, such as contact profiles and synaptic positions, can be rigorously quantified and systematically represented as graphs capturing multidimensional relationships. These representations require methodologies from data science that enable dimensionality reduction and comparisons of the architecture across scales. Yet to derive new intuitions about the spatiotemporal events leading to the architecture that shapes its function, it is necessary to integrate and compare these various representations, bridging knowledge from the cell biological events to the systems-level network relationships. NeuroSCAN is designed to achieve this integration, enabling synthesis of knowledge ranging from the abstractions of neuronal relationships in C-PHATE to the cell biological features underpinning these abstractions. We provide a case study to illustrate how integration of analyses performed in NeuroSCAN can result in new insights. First, we demonstrated the discovery process with C-PHATE representations to identify neurons that undergo changes in their contactome during development. Second, we developed 3-D representations of contact sites to analyze the local neuronal regions that were identified via DC/C-PHATE analysis. Third, we visualized and compared these representations across development to identify cell biological changes in neuronal morphologies and synaptic positions across neuron classes. Our case study demonstrates the utility NeuroSCAN to facilitate exploration of neuronal relationships, leading to new insights on structural features of the connectome and hypotheses for empirical testing.

Comparisons of NeuroSCAN to other connectomics atlases

NeuroSCAN is one of several efforts centered around interpreting the C. elegans EM datasets. Other open-source tools for data exploration in C. elegans include efforts to capture neuron morphologies and synaptic information (including integration of new connectomes across larval development), to map neurotransmitter and receptor expression, and to record whole brain functional connectivity across genotypes (Witvliet et al., 2021; Altun, Z.F. et al., 2002; Cook et al., 2019; Fenyves et al., 2020; Randi et al., 2023). NeuroSCAN was inspired by tools like NemaNode and WormWiring (Witvliet et al., 2021; Cook et al., 2019), which enable 3-D visualizations of neuronal morphologies and synaptic sites with synaptic subsets restricted to pre or postsynaptic sites. In NeuroSCAN we sought to generate and integrate information beyond the synaptic connectome to include local neuronal regions (contactome) and neuronal morphologies across available developmental vEM datasets. Contactomes represent features that have been largely overlooked in connectomic datasets, and which capture circuit structures not evident by inspecting solely synaptic relationships (Brittin et al., 2018). NeuroSCAN extends existing representations to also offer user-driven experience with choice over the visualization of specific synaptic sites, the option to search for synaptic partners, and the ability to customize the color of each synaptic representation (Figure 7). NeuroSCAN representations complement resource databases like WormAtlas, which hosts digitized electron micrographs and schematics of neuron morphologies with aggregated information on each neuron (Altun, Z.F. et al., 2002). As such, NeuroSCAN extends an existing suite of opensource resources to facilitate community wide exploration of vEM datasets.

NeuroSCAN design and future directions

NeuroSCAN code and development was intentional in its design as an open-source resource that is modular and allows integration of additional features and data structures (Cantarelli et al., 2018). It is a hypothesis-generating tool that can be equally used by educators seeking to teach neuroanatomical principles, and researchers seeking to identify changes across connectome datasets. NeuroSCAN could be integrated into emerging datasets, including developmental time-courses of cell-specific transcriptomic data that would enable further insights on the molecular events underpinning neuronal development and function– from synaptogenic processes to the logic of neurotransmitter use (Packer et al., 2019; Taylor et al., 2021; Fenyves et al., 2020) and how it sustains functional connectivity (Randi et al., 2023). Future iterations of NeuroSCAN could also include positions and relationships of neurons to non-neuronal cell types, as well as the relative networks of segmented and quantified organelles within cells. NeuroSCAN could be used to compare new datasets from genetic variants, from animals trained under specific conditions or from additional developmental datasets across embryogenesis. As such, the pipeline and design of NeuroSCAN can serve as a sandbox to examine the value of the integration of datasets in exploring representations of neuronal relationships across connectomes.

NeuroSCAN forms part of a longer tradition that has leveraged the pioneering datasets generated for C. elegans connectomes towards exploring structure-function relationships in the nervous system. While the smaller scale of the C. elegans neuropil allowed us to rigorously vet the utility of these approaches, we suggest that these same methods would be beneficial in comparative studies in neuropils of other species, including those with less stereotypically formed connectomes. We suggest that contact profiles, along with neuron morphologies and synaptic partners, can act as ‘fingerprints’ for individual neurons and neuron classes. These ‘fingerprints’ can be aligned across animals of the same species to create identities for neurons. Frameworks for systematic connectomics analysis in tractable model systems such as C. elegans are critical in laying a foundation for future analyses in other organisms with up to a billion-fold increase in neurons (Toga et al., 2012). Therefore, we envision these collective efforts akin to the foundational work from C. elegans in pioneering genomic analysis and annotations ahead of the Human Genome Project (Stein et al., 2001; Collins and Fink, 1995). We believe that further integration of datasets in platforms like NeuroSCAN would be key in determining the representations and features necessary for the interpretation and analyses of other connectomes.

Methods and materials

Lead Contact

Further information and requests can be directed to Daniel.colon-ramos@yale.edu.

Data Code and Availability

Figures in this article have been generated with NeuroSCAN (Figures 5D, Figures 6-7, Figure S2G-I, Figure S3, Figure S4, Figure S5 A-B, Figure 8, Figures S6-S12, Videos S1-S4) and CytoSHOW (Figures 1-4, Figure 5A and C, Figure S1, Figure S 5C). Data can be visualized via the viewer at NeuroSCAN.net or by downloading glTF files from NeuroSCAN and using a glTF viewer to visualize them. Additionally, the data generated for NeuroSCAN is available in .OBJ file format (and can be visualized from a local hard drive with CytoSHOW (http://neuroscan.cytoshow.org/). All excel files for Diffusion Condensation iterations and adjacency quantifications can be found in Tables S3-S13. Tutorials for NeuroSCAN are available on NeuroSCAN.net upon opening the website, within the main menu of the website (Figure S8), and in the supplementary materials (Figure S5-S12; Videos S1 and S3-S4). These tutorials generally cover the process of engaging in analysis at and across specific developmental stages by filtering the data items and adding items to viewers (Figure S10). General understanding for how to use C-PHATE to analyze neuronal relationships can be found in Figure 1, Figure 4, Figure S6, Video S1, and in our previous publication (Moyle et al., 2021). For additional information on filters and in-viewer changes to the data (colors, developmental stages, downloading data) see Figure S5, Figure S7, Figure S11, Figure S12, and Videos S3-S4. All code for website development is available at Github (https://github.com/colonramoslab/NeuroSCAN) and for information on website architecture and data model see Figures S13-S14.

Experimental Model and Subject Details

Volume electron microscopy (vEM) data and segmentation of neurons and synapses were analysed from (Witvliet et al., 2021; White et al., 1986; Brittin et al., 2018; Cook et al., 2019). We analyzed available EM datasets that were transversely sectioned and segmented (Witvliet et al., 2021; Brittin et al., 2021; White et al., 1986). We deleted the CAN neurons in the L1-L3 datasets to keep these datasets consistent with the legacy datasets L4 and Adult (N2U), which do not contain CAN neurons (as in (Moyle et al., 2021)).

Method Details

All 3-D object isosurfaces (Morphologies (Neurons), Contacts, Synapses, C-PHATE plots) were generated from segmented EM datasets using a modified version of the ImageJ 3D viewer plug-in (Schmid et al. 2010) implemented in CytoSHOW (scytoshow.org). This tool employs the marching cubes algorithm for polygon-generation. All 3-D objects are first exported as wavefront (.OBJ) files then converted to GL Transmission Format (.glTF) file format which does not distort the resolution but compacts the file information to enable faster loading times in the web-based 3-D viewer.

Pixel Threshold Distance for Adjacency Profiles and Contacts

We identified two challenges in compiling Electron Microscopy (EM) datasets for comparisons: 1) how to uniformly capture neuronal relationships based on areas of physical adjacency (contact) across datasets that have differences in volume depth and in x-y-z resolutions, and 2) how to standardize across datasets in which membrane boundaries had been called using a variety of methods, including contrast methods and segmentation methods (hand-drawn vs predicted via centroid node expansion by a shallow convolutional neural network) (Witvliet et al., 2021; Brittin et al., 2018; White et al., 1986). To address this, we first standardized the region of the neuropil across all developmental stages as in (Moyle et al., 2021). Briefly, all cell bodies were deleted, and we used the entry of the nerve ring neurons into the ventral cord as the posterior boundary landmark for the entire volume, focusing on the AIY Zone 2 (Colón-Ramos et al., 2007); slice range Table S1). Previously reported adjacency profiles used 10 pixels (or 45 nm) as the pixel threshold distance for the L4 (JSH) and Adult (N2U) datasets (Moyle et al., 2021). To account for differences in resolution (x-y axis) and in calling membrane boundaries between the L4 and Adult datasets and L1-L3 datasets, we designed a protocol to define the pixel threshold for each dataset. In short, for two cells that are in direct contact (Figure S2 D) in the manually segmented datasets (L4 and Adult), we calculated the length of overlap needed to reach from the segmented edge of one cell, across the membrane, and into the adjacent cell, when the segmented area of one cell is expanded by 45 nm (10 pixels). This results in an average overlap of 30 nm for directly contacting cells in the L4 dataset. Then, in each computationally segmented dataset (L1-L3), we empirically tested the distance (e.g. 55 nm, 60 nm, 62 nm) required to achieve a similar overlap of 30 nm in direct contact cells. That empirical number (in nm) was used for adjacency calculations and rendering of contacts. The numbers were converted from nanometers into pixels to create a pixel threshold distance for each dataset, and these are shown in Table S1. Once these corrections had been applied, we calculated the cell-to-cell adjacency scores for all cell pairs in each dataset by using the measure_adjacency algorithm from https://github.com/cabrittin/volumetric_analysis; (Brittin et al., 2018) (Tables S8-S13). Adjacency matrices were used for Diffusion condensation (Brugnone et al., 2019).

Diffusion Condensation

Diffusion condensation (DC) is a dynamic, time-inhomogeneous process designed to create a sequence of multiscale data representations by condensing information over time (Brugnone et al., 2019). The primary objective of this technique is to capture and encode meaningful abstractions from high-dimensional data, facilitating tasks such as manifold learning, denoising, clustering, and visualization. The underlying principle of diffusion condensation is to iteratively apply diffusion operators that adapt to the evolving data representation, effectively summarizing the data at multiple scales. The diffusion condensation process begins with the initialization of an initial data representation, typically the raw high-dimensional data or a preprocessed version. This initial representation is used to construct a diffusion operator, a matrix derived from a similarity matrix that reflects the local geometry of the data. The similarity metric, such as Euclidean distance or cosine similarity, plays a crucial role in defining these local relationships. Once the initial diffusion operator is established, the algorithm proceeds to the diffusion step. In this step, the diffusion operator is applied to the data, smoothing it by spreading information along the edges of the similarity graph. This operation captures the intrinsic geometry of the data while reducing noise. The specific form of the diffusion operator, such as the heat kernel or graph Laplacian, significantly impacts how information is propagated during this step. Following the diffusion step, the condensation step updates the data representation by aggregating diffused data points if the distance between them falls below a ‘merge threshold’. This step creates a more compact and abstract representation of the data. These diffusion and condensation steps are iteratively repeated. At each iteration, the diffusion operator is recomputed based on the updated diffuse data representation, ensuring that the process adapts to the evolving structure of the data. The iterations continue until a stopping criterion is met, such as convergence of the data representation to a single point. The output of the diffusion condensation process is a sequence of multiscale data representations. Each representation in this sequence captures the data at a different level of abstraction, with earlier representations preserving more detailed information and later representations providing more condensed summaries. This sequence of representations can be utilized for various tasks, including manifold learning, denoising, clustering, and visualization. By iteratively smoothing and condensing the data, diffusion condensation reveals the underlying structure of high-dimensional datasets. A detailed algorithm description is provided in Box 1 and Algorithm 1.

Diffusion Condensation

Initialization

Let X = {x₁, x₂, …, x_n} be the set of n data points in a high-dimensional space. Construct the affinity matrix A, where A_ij measures the similarity between x_i and x_j. Typically,

for a chosen scale parameter σ.

Diffusion Operator

Define the degree matrix D as a diagonal matrix where D diffusion operator D_ii = ∑_j A _ij. Construct the

which normalizes the affinity matrix.

Diffusion Step

Apply the diffusion operator to the data:

This step smooths the data, capturing the intrinsic geometry.

Condensation Step

After each diffusion step, merge data points that are within a small distance, ϵ, from each other to form a condensed representation. Specifically, data points x_i and x_j are merged if

This merging process produces a set of condensed cluster centers C = {c₁, c₂, …, c}, where each center represents the mean of merged data points.

Iteration

Repeat the diffusion and condensation steps, adjusting the parameter σ adaptively, until convergence or for a predefined number of iterations.

Algorithm 1

Diffusion Condensation

C-PHATE

C-PHATE is an extension of the PHATE technique (Moon et al., 2019) which is specifically aimed at handling and visualizing high-dimensional biological data. C-PHATE is specifically designed to handle compositional data, which are datasets where the components represent parts of a whole and are inherently constrained. It learns the intrinsic manifold of the data, effectively capturing non-linear relationships and structures that are not apparent with traditional methods like PCA or t-SNE. The C-PHATE algorithm starts by loading affinity matrices associated with specific clusterings obtained from diffusion condensation. These matrices are normalized to generate kernel matrices that emphasize the strength of connections within each cluster. The algorithm then builds a connectivity matrix by integrating these kernel matrices based on cluster assignments over multiple time points. This is achieved by first initializing the matrix with kernel matrices along its diagonal and then filling in off-diagonal blocks with transition probabilities that reflect how clusters transition from one time point to the next. Next, we apply the PHATE dimensionality reduction technique to the connectivity matrix to generate 3D embeddings of the data. These embeddings are derived from multiple iterations of diffusion condensation, capturing the geometry of the data at various levels of granularity. The resulting coordinates are saved for subsequent analysis. The final step involves visualizing the PHATE results in a 3D graphics tool, CytoSHOW (Java-based; CytoSHOW.org; https://github.com/mohler/CytoSHOW; (Moyle et al., 2021)). The results are plotted in a 3D environment, with functionality enabling rollover labels to display information about clustered cells. This requires cross-referencing output tables from the original data collection. CytoSHOW is an interactive tool that allows for assigning colors and annotations to individual neurons and clusters of interest. A detailed algorithm description is provided in Box 2 and Algorithm 2. The python code for C-PHATE allows for user specification of four numerical parameters within the command line, and we used the same set of values for all C-PHATE plots shown in this report (100, 30, 50, 1). The first two integers define the weighting of connectivity between the current condensation step t and previous steps t-1 (weighting = 100) or t-2 (30), respectively, during construction of the connectivity matrix. Values 100 and 30 consistently resulted in a series of plotted clustering trajectories that form a dome-like convergence of paths, enhancing our visual perception of relative relationships and showcasing the super clusters that constitute anatomical strata in the nerve ring neuropil (Video S1). The reproducibility of the dome shape depends on assigning two specific PHATE parameters (https://phate.readthedocs.io/en/stable/api.html) to non-default values when calling PHATE, the “t” value is set to 50; the “randomstate” value is set to 1.

C-PHATE

Given n data points, X = {x₁, x₂, …, x_n}, and the diffusion condensation output, consisting of C = {c₁, c₂, …, c} denoting the merged data points and A denoting the affinity matrix at iteration.

Kernel Matrix

For each iteration,, compute the degree matrix D, where D_ii = ∑_j A _ij. Then, normalize the affinity matrix to construct the kernel matrix:

Initial Connectivity Matrix

Initialize the connectivity matrix C_PHATE with zeros. Next, populate it with the kernel matrices,, along its diagonal, reflecting self-connections within each cluster at each time point.

Update Transition Probabilities

For each pair of adjacent time points and + 1, compute a transition probability matrix to determine how points transition between clusters and C₊₁. Each entry C _j Each entry p_ij in this matrix represents the probability of moving from cluster i at time to cluster j at time + 1.p _ij is calculated by counting the number of points moving from cluster i to cluster j and normalizing by the total number of points in cluster i at time. This can be expressed as:

Use these transition probabilities to populate the off-diagonal blocks of C_PHATE

Dimensionality Reduction

Apply the PHATE algorithm to the final connectivity matrix _PHATE to obtain the low-dimensional embedding Y:

Visualization

Visualize low-dimensional embedding Y in CytoSHOW.

Box 2: Mathematical description of C-PHATE

Algorithm 2

C-PHATE

Electron Microscopy based 3-D Models

To make 3-D models of neuron morphologies from vEM datasets, we created Image-J format regions of interest (ROIs) using published segmentation data (Witvliet et al., 2021; White et al., 1986; Brittin et al., 2021). For a given cell, the stack of all sectioned ROIs was then used to draw binary image masks as input to a customized version of the marching cubes algorithm (Schmid et al., 2010) to build and save a 3-D isosurface. All steps of this pipeline were executed within the ImageJ-based Java program, CytoSHOW (AU Duncan et al., 2019). Slightly modified versions of this workflow were also followed for: 1) generating cell-to-cell contact ROIs and 2) for generating 3-D representations of synaptic objects. To align the 3-D models from the variously oriented vEM datasets, all surfaces from a given specimen were rotated and resized to fit a consensus orientation and scale. This was achieved by applying a rotation matrix multiplication and scaling factor to all vertex coordinates in isosurfaces comprising each modeled dataset (Table S2). Each 3-D object (morphology, contact or synapse) was then exported as a Wavefront file (.OBJ) and then web-optimized by conversion to a Draco-compressed .GLTF file. Each neuron was assigned a type-specific color that is consistent across all datasets to enable facile visual comparison. All the original EM annotations that were used to create the representative 3D models in NeuroSCAN have been preserved, and can be accessed via the sister app, CytoSHOW (https://github.com/mohler/CytoSHOW; (AU Duncan et al., 2019)).

Morphologies

Neuron morphologies were linked across datasets for users to visualize changes over time. To enhance 3-D graphics performance without sacrificing gross morphologies we employed a defined amount of data reduction when building each cell-morphology object. NeuroSCAN can therefore display multiple (or even all) neurons of a specimen within a single viewer. The number of vertices for a given object was decreased by reducing 10-fold the pixel resolution of the stacked 2-D masks input into the marching cubes algorithm of CytoSHOW.

Nerve Ring

To make a simplified mesh of the overall nerve ring shape, individual neuron ROIs were fused together into a single nerve-ring-scale-stack of image masks. This was used for input to the marching cubes algorithm. The union of all overlapping enlarged neurite ROIs in a vEM section was data reduced (20-fold reduced pixel resolution). This rendered a performance-friendly outer shell of the nerve ring.

Contacts

To build 3-D representations of neuron-neuron contacts, we captured the degree of overlap when an adjacent cell outline was expanded by the specimen-specific, empirically-defined pixel threshold distance listed in Table S1 (see Figure S2). This was done for each cell outline. This expansion step employs a custom-written method in CytoSHOW that increases the scale of the adjacent outlined region by the pixel threshold distance (Table S1; Figure S2 B and E), while maintaining its congruent shape. The entire collection of captured 2-D contact overlaps (Figure S2 C and F) for each adjacent neuron pairs was then reconstructed as a single 3-D object (Figure S2 H). Contact patches shown in NeuroSCAN are largely reciprocal (e.g. if there is a AIML contact from PVQL then there will be a PVQL contact from AIML), but rarely, 2-D overlap regions may be too small to be reliably converted to 3-D isosurfaces by the marching cubes algorithm, resulting in absence of an expected reciprocal contact model within the collection. Contacts, like cell morphology models, are named to be automatically linked across time-point datasets and to facilitate user-driven visualization of changes over time.

Synapses

Synaptic positions were derived from the original datasets and segmentations, which annotate synaptic sites in the EM cross-sections (White et al., 1986; Cook et al., 2019; Witvliet et al., 2021). To represent these coordinates in the 3-D segmented neurons, we used Blocks (presynaptic sites), Spheres (postsynaptic sites) and Stars (electrical synapses). The synaptic 3-D objects were placed at the annotated coordinates (White et al., 1986; Cook et al., 2019; Witvliet et al., 2021). Additionally, the objects were scaled with the scaling factor (Table S2). Synaptic objects were named by using standard nomenclature across all datasets, as explained in Supplementary Figure 7.

We note that the L4 and Adult datasets and the L1-L3 datasets were prepared and annotated by different groups (White et al., 1986; Cook et al., 2019; Witvliet et al., 2021). Integration of these datasets reveals nanoscale disagreements in the alignment of the boundaries and synapses. Our representations reflect the original annotations by the authors. Because of these disagreements in annotations, the synapses are not linked across datasets. However, all the original EM annotations that were used to create the representative 3D models in NeuroSCAN, including the synaptic annotations, have been preserved, and can be accessed by the users via the sister app, CytoSHOW (CytoSHOW.org).

Acknowledgements

We are grateful for current and former members of the Colón-Ramos lab for their guidance and suggestions, in particular, Agustín Almoril-Porras and Malcolm Díaz García for assisting with data formatting, Patricia Chanabá-López and Andrea Cuentas-Condori for feedback on the NeuroSCAN website, Mayra Blakey for administrative roles in managing contracts for funding distribution, and Ben Clark and Milind Singh for feedback on the paper. We also thank Stephen Larson, Dario Del Piano and Zoran Sinnema (MetaCell) and Jamie Emerson (Bilte Co.) for website software development, method reporting and hosting services. We thank Brandi Mattson for editing early paper drafts. We acknowledge Ryan Christensen and Hari Shroff (Janelia Research Campus) and Patrick La Riviere (University of Chicago) for helpful discussions and guidance for the NeuroSCAN website. We thank the Research Center for Minority Institutions program, the Marine Biological Laboratories (MBL), and the Instituto de Neurobiología de la Universidad de Puerto Rico for providing meeting and brainstorming platforms. D.A.C-R. acknowledges the Whitman Fellows program at MBL for providing funding and space for discussions valuable to this work. Research in D.A.C-R. and W.A.M. labs was supported by NIH grant R24-OD016474. This work was also funded by the NIH/NINDS grant R35 NS132156-01, DP1 NS111778 and R01 NS076558–2.

Additional information

Authorship Contributions

N.L.K. Conceptualization; Data curation; Investigation; Methodology; Project Administration; Validation; Visualization; Writing-Original Draft S.E.E. Conceptualization; Data curation; Formal Analysis; Investigation; Project Administration; Software; Visualization; Writing-Original Draft D.B. Formal Analysis; Software; Writing - Original Draft M.W.M. Conceptualization; Formal Analysis; Investigation; Project Administration; Software; Writing-Review, Editing P.A.-M. Data curation; Writing-Review, Editing N.V.M. Data curation; Investigation; Writing - Review, Editing S.K. Resources; Supervision W.A.M. Conceptualization; Data curation; Formal Analysis; Funding Acquisition; Methodology; Resources; Software; Supervision; Validation; Writing-Review, editing; Corresponding Author D.A.C.-R. Conceptualization; Funding Acquisition; Resources; Supervision; Visualization; Writing - Review, editing; Corresponding Author

Competing Interests

Authors do not declare any competing interests.

Declaration of generative AI and AI-assisted technologies in the writing process. During the preparation of this work the author(s) used ChatGPT in order to improve readability. After using this tool, the author(s) reviewed and edited the content as needed and take full responsibility for the content of the published article.

Supplementary material

Supplementary Figures

Supplementary Figure S1.

Supplementary Figure S2.

Supplementary Figure S3.

Supplementary Figure S4.

Supplementary Figure S5.

Supplementary Figure S6.

Supplementary Figure S7.

Supplementary Figure S8.

Supplementary Figure S9.

Supplementary Figure S10.

Supplementary Figure S11.

Supplementary Figure S12.

Supplementary Figure S13.

Supplementary Figure S14.