ALHENA, an H&E encoder distilled from transcriptomic foundation models for multi-scale slide analysis.

Documentation | Preprint | Paper code | Weights on Hugging Face | References on Hugging Face

---

ALHENA reads an H&E section and returns two things: embeddings, per tile and per cell, and cell types on an unlabeled section. No spatial transcriptomics is needed at inference, and nothing is fitted at deployment.

Cell types come from retrieval against a labeled single-cell reference, so a different reference gives a different taxonomy. ALHENA provides prebuilt references for fifteen human tissues, built from fourteen public single-cell atlases, or you can build a custom reference.

When the inputs share a spatial frame, the embeddings and the cell types stay aligned with the slide and can be analyzed together across scales.

ALHENA is gamma Geminorum, from the Arabic al-han'ah, a brand mark placed on something to make it readable. To the eye, it appears as a single star, though it is in fact a binary system. The light we see comes from a bright A-type subgiant. Its companion has never been imaged and reveals itself only through the faint wobble it leaves in the motion of the visible star. The H&E section is that visible star, inexpensive, familiar, and collected everywhere. The spatial transcriptome is the unseen companion, physically present, shaping what we observe, yet never directly accessible at inference.

Install

ALHENA needs Python 3.12 or later, and runs with or without a GPU.

pip install alhena

Optional extras add plotting (viz), whole-slide reading (wsi) and SpatialData support (spatial): pip install "alhena[viz,wsi,spatial]".

The UNI2-h model weights are distributed separately by their authors and are not included with ALHENA. Their repository is gated: request access at MahmoodLab/UNI2-h and accept the license. Then log in and download the weights, 2.7 GB, into the Hugging Face cache, where ALHENA finds them:

hf auth login
hf download MahmoodLab/UNI2-h pytorch_model.bin

For weights stored elsewhere, set ALHENA_BACKBONE to the file. ALHENA adds its own low-rank adapters and projection heads on top of the UNI2-h weights, and downloads them from Hugging Face on first use, with the prebuilt reference of each tissue it annotates.

See it work

The demo needs no data of your own. It reads a few annotated breast H&E fields from the public PanNuke dataset, annotates their nuclei with the prebuilt breast reference, and compares the predicted cell types with the PanNuke labels. It needs the UNI2-h model weights.

alhena demo Breast

Use it on your own data

import alhena

source = alhena.io.read_slide("slide.svs")                        # an H&E slide, read tile by tile
nuclei = alhena.io.read_polygons("nuclei.geojson", tiles=source)  # nucleus outlines

encoder = alhena.load_encoder()
tiles = encoder.encode_tiles(source)                              # one embedding per tile
tiles.write_h5ad("tiles.h5ad")                                    # one row per tile, ready to analyze

cells = encoder.encode_cells(source, nuclei=nuclei)               # one pass, each tile once
cells.write_h5ad("cells.h5ad")

typer = alhena.load_celltyper(tissue="Breast")                    # loads the reference alone
result = typer.annotate(cells)                                    # retrieval, no encoder
result.to_frame()

read_slide needs the wsi extra and reads the resolution the slide declares. source can also be an array of H&E tiles, (n_tiles, height, width, 3) uint8 RGB; mpp= then declares the resolution it is already at. The nuclei come from a segmenter of your choice: ALHENA does not segment. The documentation shows how to segment a slide with sopa and StarDist. read_polygons expects the outlines in the slide's full-resolution pixels. Outlines written in microns need units="um", and outlines segmented on another image need transform=. See Coordinates. With SpatialData, alhena.spatial reads the image and the nucleus outlines from the object.

Cells are encoded once and can be annotated again, against another reference or on another annotation axis, without rerunning the image encoder: alhena.io.read_cells("cells.h5ad") reads them back with the provenance that pairs them with a reference. typer.annotate(source, nuclei=nuclei) runs both steps in one call and keeps the cells on result.cells.

Documentation

prism-oncology.github.io/ALHENA

Start with Choose your task: tile embeddings, cell embeddings, cell-type annotation, or multi-scale slide analysis. Inputs says how to provide your images and nuclei, whatever form they take.

Six tutorials cover these tasks, custom references, and annotation evaluation. The use case example applies ALHENA to two public colorectal cancer cohorts, from tissue contexts to patient-level analyses.

Contributing

Report a problem or ask a question in the issues; CONTRIBUTING.md explains how to install ALHENA for development and run its tests.

License

The ALHENA software is released under the BSD 3-Clause License. The ALHENA model weights, comprising the LoRA adapters and the projection heads used at inference, are released under the Creative Commons Attribution-NonCommercial 4.0 International License (CC BY-NC 4.0).

The precomputed single-cell reference representations are released under the CC BY-NC 4.0 license. They remain subject to the terms of their source atlases, and those derived from DISCO are provided for academic research only, in accordance with the DISCO terms of use.

Citation

If you use ALHENA, please cite the preprint:

@article{alhena,
  title   = {{ALHENA}, an {H\&E} encoder distilled from transcriptomic foundation models
             for multi-scale slide analysis},
  author  = {Giraud-Sauveur, F{\'e}licie and Blampey, Quentin and Marza, Pierre and
             Benkirane, Hakim and Vakalopoulou, Maria and Courn{\`e}de, Paul-Henry and
             Christodoulidis, Stergios},
  journal = {bioRxiv},
  year    = {2026},
  doi     = {10.64898/2026.10.20.760439}
}

Please also cite the models and datasets ALHENA builds on. The Cite page of the documentation lists them.