Predicting Single-Cell Spatial Gene Expression from H&E using Pathology Foundation Models

Predicting gene expression directly from hematoxylin and eosin (H\&E) images offers a scalable and cost-effective approach to deriving spatial molecular insights from routine histology. Pathology foundation models (FMs) provide a promising basis for this task because they learn rich visual representations from large collections of H\&E images. These models are typically pretrained using tile-level objectives, making their representations well suited to spot-level spatial transcriptomics (ST), in which each measurement aggregates expression across multiple cells. However, as ST technologies advance toward single-cell resolution, we find that these tile-level representations are not directly suited for predicting gene expression at the single-cell level. In this work, we systematically evaluate and establish domain-informed and task-specific strategies for adapting pathology FMs to single-cell spatial gene expression prediction. On a Xenium 5K dataset with over 5,000 genes, our method improves mean gene-wise Pearson correlation by 2.64$\times$ over the standard zero-shot baseline and achieves state-of-the-art performance on this task. Furthermore, we stratify performance by biologically relevant gene sets, cell types, and pathways, lending valuable insights into which gene expression signals are already well captured, which benefit the most from our single-cell adaptations, and which remain challenging to infer from histology alone.