Abstract
We investigate joint graph inference for the chemical and electrical connectomes of the Caenorhabditis elegans roundworm. The C.elegans connectomes consist of 253 non-isolated neurons with known functional attributes, and there are two types of synaptic connectomes, resulting in a pair of graphs. We formulate our joint graph inference from the perspectives of seeded graph matching and joint vertex classification. Our results suggest that connectomic inference should proceed in the joint space of the two connectomes, which has significant neuroscientific implications.
Introduction
The roundworm Caenorhabditis elegans serves as a foundational model in neuroscience, being the only organism with a fully mapped "connectome." However, this neural wiring is not a monolithic network. Instead, it comprises two distinct layers: chemical synapses, which facilitate one-way signaling, and electrical gap junctions, which allow for bidirectional communication.
Historically, these networks have been analyzed in isolation. This study challenges that convention by asking whether combining these two distinct graphs into a single mathematical framework—Joint Graph Inference—can yield deeper insights into the identity and function of individual neurons.
Data Description: The Two Faces of the Connectome
The study focuses on a subset of the hermaphrodite C. elegans connectome consisting of 253 non-isolated neurons. These neurons are categorized into three functional types: Sensory Neurons (27.96%), Interneurons (29.75%), and Motor Neurons (42.29%).
The data is represented as two synaptic graphs:
- The Chemical Connectome ($G_c$): Symmetrized and binarized for this study, with a sparsity of approximately 2.8%.
- The Electrical Connectome ($G_g$): Naturally undirected and significantly sparser, at approximately 1.3%.
As seen in the adjacency matrices, while the two networks are distinct, they exhibit a "statistically significant similar connectivity structure." They are neither identical nor independent, suggesting a complementary relationship.
Methodology
The researchers employed two primary techniques to perform inference across these dual networks:
Seeded Graph Matching (SGM)
SGM is an alignment problem. Given two graphs, the goal is to match a neuron in the chemical network to its counterpart in the electrical network. By using a small set of known "seed" neurons as anchors, the SGM algorithm identifies structural correspondences across the remaining nodes.
Joint Vertex Classification via Omnibus Embedding
The core innovation of this work is the use of an Omnibus Matrix. Instead of analyzing each graph independently, the researchers construct a large matrix containing both the chemical and electrical adjacency matrices, along with their average in the off-diagonal blocks.
Through Spectral Embedding, this massive matrix is projected into a lower-dimensional space. This allows every neuron to be represented as a point that captures information from both connection types simultaneously. A Support Vector Machine (SVM) is then used to classify these points into the three functional neuron types.
Key Results
The findings strongly support the hypothesis that "joint space" inference is superior to single-graph analysis.
- Matching Accuracy: As the number of known "seeds" increased, the accuracy of matching the rest of the network improved significantly, far exceeding random chance. This confirms that the two networks share a common underlying structural logic.
- Classification Superiority: Joint vertex classification consistently outperformed single-graph classification. For the chemical connectome, incorporating electrical data lowered the error rate in predicting neuron types. For the electrical connectome, the joint approach was particularly effective at higher embedding dimensions, where the inherent sparsity of electrical-only data typically introduces noise.
Conclusion and Neuroscientific Implications
The success of joint inference suggests that the coexistence of chemical and electrical synapses is not a biological redundancy. Instead, these two systems work in tandem to define a neuron's functional identity. From a computational perspective, this study demonstrates that when multiple views of a network are available, the most accurate inferences are drawn from their joint representation rather than their individual parts.
Summary updated