Predicting Neural Circuit Vulnerability Using Graph Neural Networks: A Study on the Caenorhabditis elegans Connectome

This project, developed by Mrunmayee Wankhede, explores the use of Graph Neural Networks (GNNs) to predict circuit vulnerability in the Caenorhabditis elegans neural connectome, comparing it with traditional centrality metric benchmarks to provide new computational tools for neuroscience research. Positioned at the intersection of machine learning and neuroscience, the project aims to capture the complex dynamic properties and high-order topological structures of neural networks through data-driven methods.

The brain is a complex network, and the connectome is a comprehensive map of all connections in the nervous system. Caenorhabditis elegans is the only multicellular organism with a complete neural connection map (approximately 302 neurons and over 7000 synapses) and can exhibit complex behaviors. Neural circuit vulnerability refers to the degree to which perturbations of nodes/edges affect system function, which is crucial for understanding neurodegenerative diseases such as Alzheimer's disease.

The core innovation of the project is the application of deep learning to connectome analysis. Traditional methods rely on metrics such as degree centrality and betweenness centrality, which are simple and intuitive but struggle to capture complex dynamics and high-order structures; GNNs can automatically learn complex mappings between structural features and vulnerability, discovering patterns that are difficult for traditional methods to identify.

Data Foundation: Based on the complete connectome of C. elegans, including 302 neurons (attributes: type, location, neurotransmitter) and over 7000 synapses, modeled as an attributed heterogeneous graph. GNN Architecture: Uses a Message Passing Neural Network (MPNN), including node embedding layer, message passing layer, readout layer, and prediction head. Vulnerability Modeling: Regression (function retention ratio) or classification (vulnerable or not) tasks, using random node deletion, targeted attacks, and edge perturbations to generate data. Comparison Methods: Compared with degree, betweenness, closeness, eigenvector centrality, and PageRank, evaluating accuracy, efficiency, and interpretability.

Advantages of GNNs: Higher prediction accuracy (identifying hidden key nodes), sensitivity to high-order structures (e.g., motifs, communities), and integration of structural and attribute information. Value of Traditional Methods: Efficient computation, strong interpretability, and serving as benchmarks. Neuroscience Implications: Identifying key circuits (disease targets), understanding selective vulnerability in diseases, and inspiring the design of artificial neural networks.

The project provides complete code, including data preprocessing, GNN implementation based on PyTorch Geometric, benchmark method calculation, evaluation framework, and visualization tools. The code has a clear structure and complete documentation, with an open-source license encouraging community contributions, making it easy to reproduce and extend.

Limitations: Small number of neurons in C. elegans (302), limited functional data, and the model does not integrate the temporal characteristics of neural dynamics. Future Directions: Dynamic GNNs (modeling temporal evolution), multimodal fusion (structure + function + genetic data), causal inference, and cross-species transfer (application to larger nervous systems).

This project demonstrates the potential of GNNs in connectome analysis, with prediction performance superior to traditional methods, providing a new framework for computational neuroscience. The value of interdisciplinary research is significant: it provides neuroscientists with tools for structure-function relationships, shows machine learning researchers the scientific applications of GNNs, and opens up research paths from model organisms to complex nervous systems.