Deciphering spatial domains from spatially resolved transcriptomics with Siamese graph autoencoder.

Area of Science:

• Spatial transcriptomics
• Computational biology
• Bioinformatics

Background:

• Cell clustering is crucial for spatial transcriptomics (ST) data analysis.
• Graph neural network (GNN) methods for ST analysis face representation collapse, limiting clustering performance.

Purpose of the Study:

• To introduce SGAE, a novel Siamese graph autoencoder framework for improved spatial domain identification in ST data.
• To overcome the limitations of existing GNN-based methods, specifically representation collapse.

Main Methods:

• Developed SGAE, a Siamese graph autoencoder framework.
• Constructed a graph integrating gene expression and spatial information from ST data.
• Mitigated information correlation at sample and feature levels to enhance representation discrimination.

Main Results:

• SGAE effectively captures spatial patterns and generates high-quality cell clusters.
• Achieved superior performance compared to alternative ST clustering methods, validated by Adjusted Rand Index, Normalized Mutual Information, and Fowlkes-Mallows Index.
• Demonstrated utility of SGAE clustering results for accurate 3D Drosophila embryonic structure identification.

Conclusions:

• SGAE offers a robust solution for spatial domain identification in ST data.
• The framework shows potential for 3D reconstruction and tissue structure investigation.
• Source code and results are publicly available for reproducibility and further research.