VOICE estimates single-cell gene activity from histology images
A University of Michigan preprint describes VOICE, trained on 23.2 million cells across 15 tissue types, which estimates per-cell gene expression from H&E images by combining cell morphology with a reference library of Xenium-measured cells.
On August 8, a University of Michigan team posted the VOICE preprint: a model that estimates gene activity in individual cells from a standard H&E histology image. The training set covered 23.2 million cells from 75 sections across 15 human tissue types.
H&E staining (hematoxylin and eosin) reveals cell shape and tissue architecture. Xenium measures the activity of a predefined gene set in individual cells on the same tissue section, recording each cell's coordinates. VOICE was trained on paired data: one cell, one image, one set of Xenium measurements. Previously, OmiCLIP linked H&E images to transcriptomic data at the tissue-region level, where signals from multiple cells are combined. In VOICE, both the image and the Xenium measurement belong to a single cell, so the model connects a cell's appearance to its molecular profile.
The model uses two estimation routes. The direct route uses the cell's shape and immediate surroundings in the H&E image to estimate gene activity. The library route searches a reference collection for similar cells from the same tissue type and averages their Xenium values. For each gene, the model learns from the training sections how much weight to assign to each route. As the authors put it: some genes are closely tied to cell morphology and can be predicted directly from the image; others are better estimated through cells whose transcriptomic profile (a set of gene activity measurements from a single cell) is already known.
For genes with no clear morphological signal, the estimate depends on the reference library, and building that library requires Xenium measurements from cells of the same tissue type.
To test generalization, the authors held five sections out of model development and built each section's library from other sections of the same tissue, so test data did not overlap with training data. Compared with three other methods, VOICE returned the best or joint-best result across all seven single-cell prediction metrics.
How useful is this kind of prediction for a new tissue type when you still need a Xenium reference library before you can start?
Open the related Eternal Search page
Sources
[1] arxiv.org
[2] pmc.ncbi.nlm.nih.gov
VOICE is a data-rich preprint with a named institutional origin, a concrete mechanism, and an honest limitation that makes for a clean Eternal Search angle: the University of Michigan has a public organization page on the platform, so the news item connects directly to a discoverable entity. The training scale (23.2M cells, 15 tissue types) and the held-out validation design make the evidence specific enough to write about without inflating the claims.