BRIDGE

BRIDGE is a pretrained multi-organ foundation model that aligns histological image features with spatial-transcriptomics (ST) gene-expression profiles, addressing a central bottleneck in computational pathology: spatial transcriptomics remains expensive and low-throughput, while H&E-stained whole-slide images are abundant and routinely collected. By learning a shared representation that ties tissue morphology to underlying molecular state, BRIDGE aims to recover gene-expression information directly from histology and to transfer that knowledge to downstream clinical tasks without per-organ fine-tuning.

The model was developed by Lequan Yu's lab at The University of Hong Kong (Liang, Zhao, Wang, Chen, Huang, and Yu) and posted to bioRxiv in May 2026. Its distinguishing design choice is breadth: rather than specializing on a single tissue type, BRIDGE is trained jointly across 13 human organs and three sequencing technologies, encouraging the model to learn morphology-to-molecule relationships that generalize across anatomical contexts and assay platforms.

This places BRIDGE among an emerging class of histology-to-expression foundation models that treat paired image and ST data as two views of the same tissue, contrasting with earlier methods trained narrowly on one organ or one platform. The central claim is that scale and diversity of paired data, rather than task-specific tuning, drive strong zero-shot transfer.

#Key Features

• Multi-organ coverage: Pretrained jointly across 13 human organs, learning histology-to-expression relationships that generalize rather than overfitting to a single tissue type.
• Cross-platform training: Spans three spatial-transcriptomics sequencing technologies, improving robustness to platform-specific batch effects and resolution differences.
• Large paired corpus: Trained on more than 600,000 paired histology-ST profiles assembled into the BIG-600K dataset, referenced via Mendeley Data.
• Zero-shot virtual ST: Predicts gene expression from histology without per-organ fine-tuning, reporting an average PCC of 0.474 on 80 biomarker genes (roughly 30% over the prior state of the art).
• Zero-shot survival prediction: Transfers to cancer prognosis with an average C-index of 0.717 across three unseen cancer types.

#Technical Details

BRIDGE couples a histology image encoder with a gene-expression representation and aligns the two modalities so that image features can be mapped to ST profiles. Training uses the BIG-600K corpus of over 600,000 paired histo-ST profiles drawn from 13 organs and three sequencing platforms. The authors report two headline zero-shot results obtained without any per-organ fine-tuning: virtual ST prediction reaching an average Pearson correlation coefficient (PCC) of 0.474 across 80 biomarker genes, described as approximately 30% above the prior state of the art, and cancer survival prediction achieving an average concordance index (C-index) of 0.717 across three unseen cancer types. These figures are from the bioRxiv preprint (v1, posted 2026-05-08) and have not yet undergone peer review.

#Applications

BRIDGE is aimed at researchers and computational pathologists who want molecular-level readouts from inexpensive, widely available histology slides. Virtual ST prediction can supplement or pre-screen costly sequencing experiments, prioritizing regions or cases for wet-lab follow-up, while the zero-shot survival capability points toward histology-driven prognostic tools that span multiple cancer types. Because the model is designed to work across organs and platforms without retraining, it is particularly relevant for settings where paired ST data for a given tissue is scarce or unavailable.

#Impact

By demonstrating that a single multi-organ, multi-platform model can deliver strong zero-shot transfer to both expression and survival tasks, BRIDGE reinforces the case for scale and data diversity in histology-to-molecule foundation models. Important caveats apply: the work is a non-peer-reviewed preprint. A public code repository is available on GitHub and pretrained checkpoints are distributed (via Google Drive), with the BIG-600K dataset released on Mendeley Data — but the repository carries no software license and the checkpoints no data license, so despite being downloadable their downstream reuse rights are unspecified (the preprint itself declares no reuse license). Independent reproduction will be needed to confirm the reported benchmark gains.