Slide-level batch structure limits histology-guided supervision of transcriptomic foundation models

Spatial transcriptomics merges spatial gene expression data with tissue morphology within a single tissue section. Transcriptomic foundation models utilize these expression profiles to generate versatile representations, although they may maintain slide- and cohort-specific variations that mask biological signals. The researchers investigated if matched H&E histology could enhance these representations as a training-time supervisory signal that is discarded during inference.

Across three transcriptomic foundation models, histology-guided supervision yielded no consistent aggregate improvement in cross-donor annotation transfer, even though it facilitated stronger transfer from histology alone. The researchers discovered that gene-expression embeddings from spot-based spatial transcriptomics are low-dimensional and heavily structured by slide identity, which limits the shared gene-morphology signal available for cross-modal transfer.

Guidance proved beneficial for some morphologically distinctive classes but diminished held-out gene predictivity broadly across the transcriptome. These findings imply that cross-modal supervision can reorganize information already encoded in a fixed representation but cannot retrieve information the representation does not retain.

This underscores the significance of identifying slide-specific structure before implementing such supervision. The code for this study is accessible at https://github.com/ratschlab/vision-guided-transcriptomics-fm-2026.