Self-Supervised Neuron Morphology Representation With Graph Transformer.

Area of Science:

• Neuroscience
• Computational Biology
• Machine Learning

Background:

• Neuronal morphology representation is crucial for brain function studies but challenged by complex intra-class variations.
• Existing methods struggle to balance robustness and discriminative power for diverse neuronal structures.

Purpose of the Study:

• To develop a robust and comprehensive method for neuronal morphology representation.
• To improve neuron classification, retrieval, and functional property prediction.

Main Methods:

• Proposed SGTMorph, a hybrid Graph Transformer framework combining graph neural networks and Transformers.
• Incorporated random walk-based positional encoding and a spatially invariant encoding mechanism.
• Utilized a self-supervised training strategy based on geometric and topological similarity.

Main Results:

• SGTMorph demonstrated superior performance in neuron morphology classification and retrieval across five datasets.
• Accurately predicted functional properties, including soma laminar distribution and axonal projection patterns.
• The framework effectively encodes neuronal structural information with biological fidelity.

Conclusions:

• SGTMorph offers a robust and adaptable solution for neuronal morphology representation.
• The method advances computational neuroscience by enabling label-free feature learning and functional prediction.
• Publicly available code facilitates broader adoption in neuroscience research.