SynapseCLR: Uncovering features of synapses in primary visual cortex through contrastive representation learning

Area of Science:

• Neuroscience
• Computational Biology
• Machine Learning

Background:

• 3D electron microscopy (EM) is generating vast brain connectomics datasets.
• Scalable analysis methods are crucial for interpreting these large-scale neural circuit visualizations.

Purpose of the Study:

• To introduce SynapseCLR, a self-supervised contrastive learning method for 3D EM data.
• To extract and analyze synaptic features from mouse visual cortex connectomics data.
• To demonstrate SynapseCLR's utility in downstream analysis tasks.

Main Methods:

• Implemented SynapseCLR, a self-supervised contrastive learning framework.
• Applied the method to 3D EM datasets of mouse visual cortex.
• Utilized learned feature representations for synapse classification and analysis.

Main Results:

• SynapseCLR effectively separates synapses based on appearance and structural annotations.
• Achieved >99.8% accuracy in assigning excitatory vs. inhibitory neuronal types to synapses and neurites.
• Enabled efficient identification of segmentation errors and annotation imputation with minimal manual input (0.2% of data).
• Revealed intrinsic axes of synaptic structural variation and inhibitory subtype information within learned representations.

Conclusions:

• SynapseCLR provides a powerful, scalable approach for analyzing 3D EM connectomics data.
• The method facilitates detailed analysis of synaptic structure and neuronal connectivity.
• Enables advanced connectomics studies, including neurite-enhanced analysis and exploration of synaptic diversity.