Neuronal classification from network connectivity via adjacency spectral embedding.

Area of Science:

• Computational Neuroscience
• Network Science
• Graph Theory

Background:

• Understanding neural circuitry is crucial for deciphering brain function.
• Classifying neurons based on connectivity is a key challenge in connectomics.
• Stochastic block models (SBM) offer a framework for analyzing network communities.

Purpose of the Study:

• To develop and validate a novel computational strategy for neuron classification using graph-based methods.
• To accurately identify neuron types within large-scale, biologically realistic neural networks.
• To assess the robustness and scalability of the proposed classification method.

Main Methods:

• Utilized adjacency spectral embedding on a stochastic block model (SBM) graph of neural connectivity.
• Employed a Gaussian mixture model-based expectation maximization (EM) clustering algorithm for neuron assignment.
• Incorporated hierarchical agglomerative clustering for improved EM initialization and multiple restarts for enhanced accuracy.

Main Results:

• The method successfully classified eight neuron classes in a large-scale (2^12-2^15 neurons) neural network.
• The approach demonstrated stability across different embedding dimensions and excellent scalability with increasing network size.
• Clustering accuracy remained high, achieving perfect classification even with up to 40% simulated experimental noise.

Conclusions:

• The proposed strategy provides a robust and scalable method for neuron classification based on network circuitry.
• This approach holds significant potential for analyzing and interpreting complex, large-scale brain connectomics data.
• The findings contribute to a better understanding of underlying cellular components within neural networks.