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

Abstract

Spatial transcriptomics pairs spatially resolved gene expression with tissue morphology in the same tissue section. Transcriptomic foundation models encode such expression profiles into general-purpose representations, but these representations can retain slide- and cohort-specific variation that obscures biological signal. Here, we test whether matched H\&E histology can improve these representations as a training-time supervisory signal that is discarded at inference. Across three transcriptomic foundation models, histology-guided supervision provides no consistent aggregate improvement in cross-donor annotation transfer, despite substantially stronger transfer from histology alone. We find that gene-expression embeddings from spot-based spatial transcriptomics are low-dimensional and strongly structured by slide identity, limiting the shared gene--morphology signal available for cross-modal transfer. Guidance benefits some morphologically distinctive classes but reduces held-out gene predictivity broadly across the transcriptome. These results suggest that cross-modal supervision can reorganize information already encoded in a frozen representation but cannot recover information the representation does not retain, highlighting the importance of diagnosing slide-specific structure before applying such supervision. Code availability: https://github.com/ratschlab/vision-guided-transcriptomics-fm-2026

Spatially guided translation from histology images to transcriptomic profiles using foundation model-driven contrastive learning
Zi Huai Huang, Ziyang Xu, Pingzhao Hu
PLOS Computational Biology8 Sep 2026Peer-reviewedCC BY 4.0Full textBioinformatics & computational biology Medical imaging & radiologyPLOS
A multimodal foundation model to integrate histology with spatial molecular programs
Z. Zhang, B. Qin, Y. Zhao +25
bioRxiv1 Jun 2026PreprintCC BY-NCFull text1 citedBioinformatics & computational biologybioRxiv
ALHENA, an H&E encoder distilled from transcriptomic foundation models for multi-scale slide analysis
Félicie Giraud-Sauveur, Quentin Blampey, Pierre Marza +4
HAL20 Oct 2026PreprintCC BY 4.0Full textBioinformatics & computational biologyHALbioRxiv
LIMBER: A Hybrid RNA Foundation Model for Predicting and Understanding RNA Structure and Function
Z. Wang, Y. Qiu, Y. He +4
bioRxiv28 Oct 2026PreprintCC BY-NC-NDFull textBioinformatics & computational biologybioRxiv
Subcellularly Resolved Single-Cell Embedding Learning with Transcriptomic data, Protein Structure and Localization Information
Zhen Zhou, Jiachen Li, Yuan Liu +2
arXiv2 Sep 2026PreprintarXiv licenceFull textGenetics & genomicsarXiv

Cite

BibTeX RIS

BibTeX

@misc{mcconnell2026slidelevel,
  title = {{Slide-level batch structure limits histology-guided supervision of transcriptomic foundation models}},
  author = {U. McConnell and K. Nonchev and V. H. Koelzer and G. Raetsch},
  howpublished = {bioRxiv},
  year = {2026},
  month = sep,
  doi = {10.64898/2026.09.30.755585},
  url = {https://www.biorxiv.org/content/10.64898/2026.09.30.755585v1},
  note = {CC BY},
}

RIS

TY  - UNPB
TI  - Slide-level batch structure limits histology-guided supervision of transcriptomic foundation models
AU  - U. McConnell
AU  - K. Nonchev
AU  - V. H. Koelzer
AU  - G. Raetsch
PY  - 2026
DA  - 2026/09/30
PB  - bioRxiv
DO  - 10.64898/2026.09.30.755585
UR  - https://www.biorxiv.org/content/10.64898/2026.09.30.755585v1
AB  - Spatial transcriptomics pairs spatially resolved gene expression with tissue morphology in the same tissue section. Transcriptomic foundation models encode such expression profiles into general-purpose representations, but these representations can retain slide- and cohort-specific variation that obscures biological signal. Here, we test whether matched H\&E histology can improve these representations as a training-time supervisory signal that is discarded at inference. Across three transcriptomic foundation models, histology-guided supervision provides no consistent aggregate improvement in cross-donor annotation transfer, despite substantially stronger transfer from histology alone. We find that gene-expression embeddings from spot-based spatial transcriptomics are low-dimensional and strongly structured by slide identity, limiting the shared gene--morphology signal available for cross-modal transfer. Guidance benefits some morphologically distinctive classes but reduces held-out gene predictivity broadly across the transcriptome. These results suggest that cross-modal supervision can reorganize information already encoded in a frozen representation but cannot recover information the representation does not retain, highlighting the importance of diagnosing slide-specific structure before applying such supervision. Code availability: https://github.com/ratschlab/vision-guided-transcriptomics-fm-2026
ER  -

Abstract

Related

Cite