stDiff: a diffusion model for imputing spatial transcriptomics through single-cell transcriptomics

Area of Science:

• Genomics
• Computational Biology
• Bioinformatics

Background:

• Spatial transcriptomics (ST) reveals gene expression in tissues but faces limitations in gene detection and imaging.
• Current ST enhancement methods often rely on scRNA-seq cell similarity.
• Imaging-based methods offer high resolution but are restricted in gene number or detection sensitivity.

Purpose of the Study:

• To introduce stDiff, a novel method for enhancing spatial transcriptomics data.
• To leverage gene expression abundance relationships in scRNA-seq data for ST enhancement.
• To improve cell population identification and spatial pattern reconstruction in ST data.

Main Methods:

• stDiff utilizes a conditional diffusion model based on scRNA-seq data.
• The model employs two Markov processes: one for noise introduction and one for denoising.
• Original ST data is integrated into the denoising process to predict missing information.

Main Results:

• stDiff demonstrated exceptional preservation of cell topological structures across 16 datasets.
• The model accurately reconstructed diverse spatial expression patterns and delineated spatial boundaries.
• Enhancement outcomes closely mirrored actual ST data, unifying observed and predicted segments.

Conclusions:

• stDiff offers a robust solution for cell population identification in spatial transcriptomics.
• The method effectively enhances ST data by leveraging scRNA-seq gene expression relationships.
• stDiff is poised to advance spatial transcriptomics imputation methodologies.