Spatial domain detection using contrastive self-supervised learning for spatial multi-omics technologies.

Area of Science:

• Computational biology
• Spatial transcriptomics
• Bioinformatics

Background:

• Spatially-resolved omics technologies are advancing, necessitating computational tools for spatial domain prediction.
• Histological images and immunofluorescence (IF) staining offer complementary views of tissue architecture.

Purpose of the Study:

• To introduce Proust, a scalable computational tool for predicting discrete spatial domains from multi-omics data.
• To leverage graph-based contrastive self-supervised learning for low-dimensional biological profile representation.

Main Methods:

• Proust integrates multiple data modalities: spatial multi-omics (RNA, protein) and histology images (H&E).
• It employs graph-based contrastive self-supervised learning to create low-dimensional representations of biological profiles.
• The method is designed for scalability to handle large spatial datasets.

Main Results:

• Proust accurately predicts discrete spatial domains within tissue samples.
• Integration of multiple data modalities consistently enhances domain detection accuracy.
• Performance was validated across diverse benchmark datasets and technological platforms.

Conclusions:

• Proust offers a robust and scalable solution for spatial domain prediction using integrated multi-omics data.
• The tool enhances understanding of tissue architecture by combining diverse biological information.
• Its accuracy and versatility make it valuable for spatial biology research.