A Benchmark Analysis of Graph and Non-Graph Methods for Caenorhabditis Elegans Neuron Classification

A research team led by Jingqi Lu and Keqi Han has established the first comprehensive benchmark for classifying neurons in the C. elegans worm, comparing modern graph neural networks against traditional machine learning approaches. Their study, published on arXiv, systematically evaluated four graph methods (GCN, GraphSAGE, GAT, and GraphTransformer) against four non-graph baselines (Logistic Regression, MLP, LOLCAT, and NeuPRINT) using the worm's complete functional connectome. The researchers classified neurons into three fundamental types—Sensory, Interneuron, and Motor—based on three distinct feature sets: Spatial (physical location), Connection (neural wiring patterns), and Neuronal Activity (temporal firing data).

The results revealed that attention-based GNNs, particularly Graph Attention Networks (GAT) and Graph Transformers, significantly outperformed all baseline methods when using Spatial and Connection features, achieving superior accuracy in neuron type prediction. However, performance using Neuronal Activity features was poor across all models, which the authors attribute to the low temporal resolution of the underlying calcium imaging data. This benchmark validates graph neural networks as powerful tools for analyzing biological networks where relationships and structure are paramount, and it highlights that spatial arrangement and connection topology are the key predictive features for neuron classification in C. elegans. The publicly released code provides a foundation for future work at the intersection of computational neuroscience and graph AI.