Compositional and interpretable representation of histology using AI foundation models and sparse autoencoders

Table of Contents

• General Information
  • Associated Publication
  • Recommended Citation
  • Useful Links
• Demo for the FM-SAE workflow
  • Installation
  • Demo notebook
• Additional notes / comments

General Information

Compositional and interpretable representation of histology using AI foundation models and sparse autoencoders

Authors: Ziyuan Zhao*, Zoltan Maliga*, Emmanuel C. Ogbonna, Soheil R. Talemi, Shannon Coy, Andréanne Gagné, Kapongo Lumamba, Isaac H. Solomon, Sandro Santagata, Adrie J.C. Steyn, Threnesan Naidoo†, Peter K. Sorger†

*Co-first Authors: Z.Z., Z.M.

†Co-Senior Authors: T.N., P.K.S.
​
Please cite this data as the following:
Zhao, Z. et al. (2026). Compositional and interpretable representation of histology using AI foundation models and sparse autoencoders. (biorxiv)

Relevant links:

• Publication DOI: https://doi.org/10.64898/2026.06.03.725182
• Associated GitHub Repository: https://github.com/labsyspharm/Zhao-FMSAE-2026
• To view an archived record of this repository: https://zenodo.org/records/21796522 ​

Licenses/restrictions placed on the data: CC-BY creativecommons.org/licenses/by/4.0/

Applying the FM-SAE workflow

Installation

Make a minimal working environment for running the demo notebook:

conda create -n Zhao-FMSAE-2026 python=3.10 -y
conda activate Zhao-FMSAE-2026
pip install .

The notebook shows how to apply a pre-trained SAE model to patch-level FM embeddings from UNI and then make multi-feature maps for whole-slide image outside of the training dataset. You can use the provided SAE model on other H&E datasets of your choice as long as you use UNI to embed patches with size of ~120 x 120 µm.

Data availability

You can find the full datasets for DS1-5 used in the paper on Zenodo: https://zenodo.org/records/21299341.

The data can be downloaded from Amazon S3 buckets following instructions on that page.