Reference-informed spatial domain detection using weak supervision for spatial transcriptomics

Area of Science:

• Genomics
• Computational Biology
• Bioinformatics

Background:

• Spatial transcriptomics (ST) aims to map tissue organization and function.
• Accurate tissue segmentation is crucial for ST data analysis.
• Existing methods often require extensive annotations or lack generalizability.

Purpose of the Study:

• To introduce GraphScrDom, a novel reference-informed, weakly supervised contrastive learning model.
• To integrate manual annotations (scribbles) with single-cell RNA-seq data for tissue segmentation.
• To develop a user-friendly toolkit for spatial domain detection.

Main Methods:

• GraphScrDom utilizes contrastive learning to integrate spatial grid/histology image annotations with cell type-specific gene expression profiles.
• The model is reference-informed and weakly supervised.
• An integrative software toolkit with an interactive annotation interface and model training module was developed.

Main Results:

• GraphScrDom consistently outperforms existing methods in tissue segmentation across various ST platforms.
• The model demonstrates strong generalizability and robustness at both bulk and single-cell resolutions.
• Performance was validated using six widely adopted metrics.

Conclusions:

• GraphScrDom provides a robust and efficient solution for spatial domain detection in ST studies.
• The integrated toolkit facilitates user-friendly spatial domain analysis.
• This approach enhances the mapping of complex tissue organization and function.