MicNet: integrating spatially resolved transcriptomes and pathology images by contrastive deep neural network

Area of Science:

• Computational Biology
• Bioinformatics
• Genomics

Background:

• Spatially resolved transcriptomic technologies offer unprecedented insights into cellular function within tissue microenvironments.
• Integrating diverse omics data with high-resolution imaging is crucial for comprehensive biological understanding but presents significant computational challenges.

Purpose of the Study:

• To develop a novel unsupervised representation learning method, MicNet, for integrating spatial transcriptomic data and pathology images.
• To create a shared feature space enabling joint biological interpretation of molecular and spatial information.

Main Methods:

• MicNet employs an unsupervised representation learning framework to project pathology images and transcriptomic data into a common domain.
• The method maximizes feature correlation within samples while minimizing it across different samples, facilitating data integration.

Main Results:

• MicNet demonstrated superior performance compared to existing methods in key spatial analysis tasks.
• The approach successfully enabled accurate spatial domain detection, identification of spatially variable genes, and visualization of cellular organization.

Conclusions:

• MicNet provides a powerful new tool for integrating spatial transcriptomics and pathology imaging data.
• This method advances the biological interpretation of complex tissue architectures and cellular interactions.