SAGE: Spatially Aware Gene Selection and Dual-View Embedding Fusion for Domain Identification in Spatial...

Area of Science:

• Genomics
• Computational Biology
• Bioinformatics

Background:

• Spatial transcriptomics (ST) enables high-resolution gene expression mapping but struggles with accurate domain segmentation due to complex tissue structures and current method limitations.
• Existing ST analysis methods often rely on local information, lack gene interpretability, and fail to capture long-range functional relationships, hindering the resolution of biologically meaningful architectures.

Purpose of the Study:

• To present Spatially Aware Gene selection and dual-view Embedding fusion (SAGE), a unified framework for improved domain identification in spatial transcriptomics.
• To address limitations in current ST methods by integrating topic-driven gene selection and dual-view embedding fusion for enhanced spatial domain delineation.

Main Methods:

• SAGE integrates non-negative matrix factorization (NMF)-based topic modeling with classifier-based importance scoring to select spatially informative genes.
• It fuses local expression graphs with topic-driven non-local graphs using consensus refinement and contrastive graph representation learning to jointly learn spatial and functional embeddings.

Main Results:

• SAGE outperforms existing methods in clustering accuracy across 34 real-world ST datasets.
• The framework reveals functionally coherent regions, interpretable gene expression patterns, and spatial heterogeneity in human breast cancer.
• SAGE refines tumor-muscle interfaces in zebrafish melanoma and uncovers shared vascular signatures between anatomically separate tissues.

Conclusions:

• SAGE provides accurate spatial domain delineation across diverse ST platforms.
• The framework is effective for dissecting microenvironmental niches and understanding long-range tissue interactions relevant to disease progression.