Learning meaningful representation of single-neuron morphology via large-scale pre-training

Area of Science:

• Neuroscience
• Computational Biology
• Bioinformatics

Background:

• Single-neuron morphology is crucial for understanding neuronal function, development, aging, and disease.
• Existing computational methods struggle with the scale and accuracy of modern neuron morphology reconstructions.

Purpose of the Study:

• To develop MorphRep, a novel model for learning effective representations of neuron morphology.
• To overcome limitations in scale and accuracy of current computational approaches.

Main Methods:

• Utilized graph transformers to encode neuron morphology data.
• Pre-trained MorphRep on over 250,000 neuron morphology reconstructions.
• Employed data augmentation and consistency enforcement for robust feature learning.

Main Results:

• Achieved state-of-the-art performance on benchmark datasets.
• Accurately characterized neuron morphology across morphometrics, cell types, brain regions, and ages.
• Demonstrated ability to distinguish neurons under diverse conditions like genetic perturbation, disease, and environmental changes.

Conclusions:

• MorphRep offers an effective strategy for embedding and representing neuron morphology.
• Provides a valuable tool for integrating cell morphology into single-cell multiomics analysis.
• Enables deeper insights into neuronal structure-function relationships and disease mechanisms.