STCGAN: a novel cycle-consistent generative adversarial network for spatial transcriptomics cellular deconvolution

Area of Science:

• Genomics
• Computational Biology
• Bioinformatics

Background:

• Spatial transcriptomics (ST) offers insights into tissue architecture and cellular heterogeneity.
• Accurate cell-type deconvolution from sparse ST data is crucial but challenging.
• Existing methods often fail to capture tissue complexity at the single-cell level.

Purpose of the Study:

• To develop a novel computational method for accurate cell-type deconvolution in spatial transcriptomics.
• To improve the reconstruction of cell spatial distribution within tissues.
• To address limitations of current methods in capturing single-cell level tissue complexity.

Main Methods:

• Propose STCGAN, a cycle-consistent generative adversarial network (CGAN) for spatial transcriptomic data.
• Utilize CGAN pre-training for robust latent representations and consistent data mapping.
• Integrate single-cell RNA sequencing (scRNA-seq) with ST data using a trainable cell-to-spot mapping matrix.
• Incorporate spatial-aware regularization to enhance cellular distribution reconstruction.

Main Results:

• STCGAN accurately estimates cellular composition within spatial transcriptomic spots.
• The method effectively reconstructs the spatial distribution of cells across tissues.
• Benchmarking demonstrates superior cell-type deconvolution performance compared to seven state-of-the-art methods on diverse datasets.

Conclusions:

• STCGAN provides a significant advancement in spatial transcriptomic data analysis.
• The method enhances the understanding of tissue architecture and cellular heterogeneity.
• STCGAN offers a robust solution for single-cell level deconvolution and spatial reconstruction.