Diffusion-based representation integration for foundation models improves spatial transcriptomics analysis.

Area of Science:

• Computational Biology
• Genomics
• Bioinformatics

Background:

• Spatial transcriptomics (ST) data captures gene expression with spatial context, enabling analyses like cell-type annotation and clustering.
• Existing foundation models excel on single-cell RNA sequencing (scRNA-seq) but often neglect spatial information.
• There's a need for foundation models that can effectively leverage ST data for generalizable embeddings across tasks.

Purpose of the Study:

• To propose DRIFT, a novel framework integrating spatial context into foundation models for ST data.
• To enhance the generalizability and performance of foundation models on ST analysis tasks.
• To bridge the gap between scRNA-seq foundation models and the unique requirements of ST data.

Main Methods:

• Leveraging diffusion on spatial graphs derived from ST data.
• Utilizing heat kernel diffusion to propagate embeddings across spatial neighborhoods.
• Integrating local neighborhood context with transcriptomic representations from single-cell foundation models.

Main Results:

• Systematic benchmarking of five foundation models (scRNA-seq and ST-based) across key ST tasks.
• Demonstrated significant performance improvements of DRIFT on ST data compared to existing methods.
• DRIFT enhances the capabilities of foundation models for annotation, alignment, and clustering tasks.

Conclusions:

• DRIFT is an effective, accessible, and generalizable framework for modeling ST data.
• The framework successfully integrates spatial context, improving foundation model performance on ST tasks.
• DRIFT advances the development of universal models for spatial transcriptomics analysis.