Boundary-Aware Clustering of Spatial Transcriptomics Data Via Fourier Feature Mapping and Dynamic Self-Supervision

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Spatial transcriptomics data is crucial for understanding tissue architecture and cellular interactions.
• Spatial heterogeneity and gene expression discontinuities challenge precise identification of fine-grained tissue domains.
• Existing clustering methods struggle with ambiguous boundaries in spatial transcriptomic data.

Purpose of the Study:

• To develop a novel computational method for accurate identification of fine-grained spatial domains in tissues.
• To address the challenge of ambiguous spatial boundaries caused by data heterogeneity in spatial transcriptomics.
• To improve the delineation of complex tissue structures using advanced graph convolutional networks.

Main Methods:

• Proposed FPS-MGCN, a multi-view graph convolutional network integrating Fourier feature mapping and dynamic pseudo-label self-supervision.
• Introduced Fourier sine-cosine mapping to enhance sensitivity to high-frequency spatial gradients and tissue boundaries.
• Implemented a dynamic pseudo-label self-supervision strategy for progressive semantic constraints and noise reduction.

Main Results:

• FPS-MGCN successfully identified critical fine-grained structures often missed by traditional methods.
• The model demonstrated superior performance in spatial domain identification compared to state-of-the-art approaches.
• Enhanced recognition in spatial boundary regions was achieved through Fourier mapping and self-supervision.

Conclusions:

• FPS-MGCN offers a robust solution for precise spatial domain identification in complex tissues.
• The integration of Fourier features and self-supervision effectively handles spatial heterogeneity.
• This method advances the analysis of spatial transcriptomics data for biological discovery.