SpatialFusion: A Unified Model for Integrating Spatial Transcriptomics to Unveil Cell-type Distribution, Interaction,...

Area of Science:

• Genomics
• Computational Biology
• Bioinformatics

Background:

• Spatial transcriptomics (ST) advances tissue structure and intercellular interaction studies.
• Current ST methods struggle with accurate, robust, and efficient spatial domain identification and cell type deconvolution.

Purpose of the Study:

• Introduce SpatialFusion, a deep learning model to enhance spatial domain identification and cell type deconvolution.
• Integrate gene expression and spatial coordinates for improved ST analysis.

Main Methods:

• Utilize graph neural networks (GNN) and attention mechanisms for spatial relationship capture.
• Employ multi-dimensional embeddings and a dual-encoding strategy (co-learning of spatial graphs and feature maps).
• Incorporate self-supervised contrastive learning to boost accuracy and robustness.

Main Results:

• SpatialFusion outperforms existing methods in accuracy and resolution on the human DLPFC dataset.
• The model accurately maps spatial cell type distributions, showing robustness to noise and low cell density.
• Analysis of breast cancer tumor microenvironment revealed spatial heterogeneity and potential therapeutic targets (COX6C, CCND1).

Conclusions:

• SpatialFusion offers a significant advancement in spatial transcriptomics analysis.
• The model provides valuable insights for precision medicine, particularly in cancer research.
• SpatialFusion enhances understanding of tissue architecture and cellular composition.