NeuroGT: Biophysically grounded graph transformers for self-supervised representation learning of neuronal morphology

Area of Science:

• Computational Neuroscience
• Machine Learning
• Neuroimaging

Background:

• Neuronal morphology is key to brain function, but analyzing large-scale reconstructions is challenging.
• Existing methods lack biophysical grounding and struggle to balance local geometry with global structure.
• Self-supervised learning often uses single objectives, limiting holistic understanding of neuron structure-function relationships.

Purpose of the Study:

• To develop a computationally efficient and biologically faithful representation framework for neuronal reconstructions.
• To integrate biophysical inductive biases into a self-supervised learning paradigm for improved neuronal analysis.
• To overcome limitations of existing methods in capturing both local and global neuronal morphology.

Main Methods:

• Introduced NeuroGT, a Graph Transformer framework with a hybrid self-supervised learning approach.
• Developed Electrotonic Positional Encoding (EPE) and k-means Shortest Path Encoding (k-SPE) for domain-specific structural knowledge.
• Implemented a multi-task objective combining graph-level contrastive learning and node-level coordinate denoising.

Main Results:

• NeuroGT achieved state-of-the-art performance on cell-type classification and neuron retrieval tasks.
• The learned embeddings demonstrated alignment with anatomical brain-region organization.
• The framework revealed significant morphology-function relationships, enhancing neurobiological discovery.

Conclusions:

• NeuroGT provides a powerful tool for learning biologically faithful and interpretable neuronal representations.
• The integrated biophysical biases and multi-task learning effectively capture complex neuronal structures.
• This approach bridges data-driven representation learning with neurobiological insights for deciphering brain function.