Identifying batch-integrated domains from spatial transcriptomics via graph autoencoder with contrastive learning...

Area of Science:

• Computational Biology
• Genomics
• Bioinformatics

Background:

• Spatially resolved transcriptomics (SRT) offers gene expression profiling with spatial context but faces limitations in histology integration and cross-modal analysis.
• Current computational methods struggle to fully leverage histology information and employ effective contrastive strategies for biological insights.

Purpose of the Study:

• To introduce GCAST, a graph contrastive autoencoder framework for integrating multimodal SRT data.
• To enhance the biological interpretability and analytical capabilities of spatial transcriptomics data by leveraging histology information.

Main Methods:

• GCAST utilizes a self-supervised strategy for deriving representations from histology images.
• The framework employs dual graph views with data augmentation and a novel contrastive learning approach incorporating histology-weighted and gene-weighted features.
• Block-diagonal graph construction is used for automatic multi-dataset alignment and batch-effect correction.

Main Results:

• GCAST effectively captures spatial gene expression patterns for tissue domain identification.
• The framework demonstrates adaptability to datasets with or without histological images.
• GCAST supports the integration of multiple datasets for joint analyses, improving batch-effect correction.

Conclusions:

• GCAST provides a unified, biologically informed framework for spatial transcriptomics analysis.
• The method facilitates deeper insights into tissue architecture and gene expression patterns.
• GCAST has the potential to advance the field of spatial transcriptomics by overcoming existing computational challenges.