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

We propose DRIFT, a framework that integrates spatial context into the input representations for foundation models by leveraging diffusion on spatial graphs derived from spatial transcriptomics (ST) data. ST captures gene expression profiles while preserving spatial context, enabling downstream analysis tasks such as cell-type annotation, clustering, and cross-sample alignment. However, due to its emerging nature, there are very few foundation models that can utilize ST data to generate embeddings generalizable across multiple tasks. Meanwhile, well-documented foundational models trained on large-scale single-cell gene expression (scRNA-seq) data have demonstrated generalizable performance across scRNA-seq assays, tissues, and tasks; however, they do not leverage the spatial information in ST data. We use heat kernel diffusion to propagate embeddings across spatial neighborhoods, incorporating the local neighborhood context of the ST data while preserving the transcriptomic representations learned by state-of-the-art single-cell foundation models.

We systematically benchmark five foundational models (both scRNA-seq and ST-based) across key ST tasks such as annotation, alignment, and clustering, ensuring a comprehensive evaluation of our proposed framework. Our results show that DRIFT significantly improves the performance of existing foundational models on ST data over specialized state-of-the-art methods. Overall, DRIFT is an effective, accessible, and generalizable framework that bridges the gap toward universal models for modeling spatial transcriptomics.