Spatially informed reference-free cell-type deconvolution for spatial transcriptomics with SpatialCD

Area of Science:

• Genomics
• Computational Biology
• Bioinformatics

Background:

• Cell-type deconvolution is crucial for analyzing spatial transcriptomics (ST) data and understanding tissue heterogeneity.
• Reference-based methods for ST deconvolution often require matched single-cell RNA sequencing data, posing practical limitations.
• Existing reference-free methods neglect spatial information, despite adjacent tissue regions often sharing similar cellular compositions.

Purpose of the Study:

• To develop a novel, spatially informed, reference-free deconvolution method for spatial transcriptomics data.
• To address the limitations of existing reference-free methods by incorporating spatial context.
• To improve the accuracy of cell-type proportion and gene expression profile estimation in ST data.

Main Methods:

• Proposed SpatialCD, a reference-free deconvolution method extending Latent Dirichlet Allocation (LDA).
• Incorporated spatial regularization to encourage neighboring spots to share similar cell-type structures.
• Validated the method on simulated data and diverse real ST datasets (MERFISH, MOB, 10× Visium, DBiT-seq).

Main Results:

• SpatialCD demonstrated improved performance over existing reference-free methods across various datasets.
• The method accurately recovered transcriptional patterns and revealed biologically coherent spatial organization.
• Achieved enhanced resolution in identifying subtle anatomical layers and region-specific cell populations in normal and diseased tissues.

Conclusions:

• SpatialCD advances statistical tools for spatial transcriptomics analysis.
• The method provides a robust reference-free approach for deconvolution, overcoming limitations of existing techniques.
• SpatialCD enriches the methodological toolkit for analyzing complex spatial gene expression patterns.