CellBridge

This repository contains the adapted model from the paper Fine-Grained and Reliable Cell Classification from Pathology Images.

CellBridge is a multi-modal contrastive learning framework for fine-grained, reliable cell classification from H&E images, pretrained on 16 million paired H&E and Xenium spatial transcriptomic profiles.

Cross-modal alignment adapts a pathology encoder to learn from frozen spatial transcriptomic embeddings. Cell-Ontology-Fallback (COF) uses branch-calibrated conformal prediction to select the supported cell-type granularity.

CellBridge requires only H&E images at inference, returning fine-grained labels when supported and broader ancestor categories when finer distinctions require molecular measurements.

• Paper: https://www.biorxiv.org/content/10.64898/2026.10.20.760347v1
• Code: https://github.com/Lotfollahi-lab/cellbridge

If this repository is helpful to you, please consider citing the associated paper:

@article {ogut2026,
    author = {Ogut, Sevda and Merchant, Arpit and Vahidi, Amirhossein and Jeong, Dabin and Birk, Sebastian and Boxall, Adam and Miraki Feriz, Adib and Sallese, Marta Rosa and Chang, Yinshui and Frossard, Pascal and Thanou, Dorina and Lotfollahi, Mo},
    title = {Fine-Grained and Reliable Cell Classification from Pathology Images},
    year = {2026},
    journal = {bioRxiv}
}