AugGCL: Multimodal graph learning for spatial transcriptomics analysis with enhanced gene and morphological data.

Area of Science:

• Genomics
• Computational Biology
• Bioinformatics

Background:

• Spatial transcriptomics offers gene expression insights in intact tissues.
• Reconstructing accurate spatial domains is challenging due to expression sparsity, complex tissue architecture, and weak signals.
• Existing methods using clustering and smoothing underperform at boundaries and in sparse regions, neglecting morphology.

Purpose of the Study:

• To introduce AugGCL, an augmented graph-convolutional learning framework.
• To enhance spatial structure decoding and gene expression reconstruction in spatial transcriptomics.
• To address limitations of traditional pipelines by integrating gene and image data.

Main Methods:

• AugGCL employs a neighborhood information aggregation mechanism integrating expression similarity and spatial proximity.
• A weighted graph and enhanced expression matrix are constructed to address sparsity without losing boundary clarity.
• A two-stream weighted graph convolutional network jointly models gene features and image-derived morphological information, using image-aware auxiliary reconstructions.

Main Results:

• AugGCL outperforms baseline methods on human prefrontal cortex, breast cancer, and mouse embryo datasets across multiple metrics.
• The method demonstrates robustness and generalization across diverse datasets.
• Downstream analyses confirm reliability for cell annotation, functional enrichment, and mechanistic studies.

Conclusions:

• AugGCL generates clearer spatial domains, advancing spatial transcriptomics applications.
• The framework effectively enhances weak spatial signals and sharpens boundaries.
• AugGCL significantly contributes to tissue structure and disease research using spatial transcriptomics.