Diffusion-based Representation Integration for Foundation Models Improves Spatial Transcriptomics Analysis

Area of Science:

• Computational Biology
• Genomics
• Bioinformatics

Background:

• Spatial transcriptomics (ST) preserves gene expression and spatial context.
• Existing foundation models for single-cell RNA sequencing (scRNA-seq) lack spatial information.
• Few foundation models are optimized for ST data, limiting generalizability across tasks.

Purpose of the Study:

• To propose DRIFT, a framework integrating spatial information into single-cell foundation models.
• To leverage spatial graphs from ST data and heat kernel diffusion for enhanced embeddings.
• To improve the performance of foundation models on ST data analysis tasks.

Main Methods:

• Developed DRIFT framework using spatial graphs from ST data.
• Applied heat kernel diffusion to propagate embeddings across spatial neighborhoods.
• Benchmarked five foundation models (scRNA-seq and ST-based) on ST tasks: annotation, alignment, and clustering.

Main Results:

• Spatial diffusion significantly improved the performance of existing single-cell foundation models on ST data.
• DRIFT outperformed specialized state-of-the-art methods for ST data analysis.
• Demonstrated enhanced performance in cell-type annotation, clustering, and cross-sample alignment.

Conclusions:

• DRIFT is an effective and generalizable framework for modeling spatial transcriptomics.
• The framework bridges the gap towards universal models for single-cell analysis.
• Spatial diffusion enhances the utility of foundation models for ST data.