VOICE teaches a model to reconstruct single-cell gene expression from a standard histology image

The VOICE model estimates gene activity in individual cells using a tissue image and a reference library of previously measured cells

On August 8, a University of Michigan team released the VOICE preprint. The authors trained the model on 23,2 million cells from 75 human tissue sections representing 15 tissues. Each cell was paired with a standard stained tissue image and gene activity measurements obtained using Xenium. VOICE extracts cell shape from the image. For genes that cannot be inferred from shape, it consults a library of similar cells whose gene activity has already been measured.

Hematoxylin and eosin staining, or H&E, shows cell morphology and tissue architecture. Xenium measures the activity of a predefined set of genes in individual cells from the same section and records their coordinates. VOICE was trained on sections in which these two data types had already been matched.

The model uses two routes to produce a single estimate. The direct route uses features of the H&E cell and its immediate surroundings to estimate gene activity. The second route searches the library for similar cells from the same tissue and averages their measured Xenium values. For each gene, VOICE uses the training sections to determine how much weight to assign to each route.

The authors describe this distinction as follows:

“Some genes are closely linked to cell morphology and can be predicted directly from the image; others are better estimated using cells with similar transcriptomic profiles.”

A transcriptomic profile is a set of gene activity measurements from a single cell. OmiCLIP had already linked H&E images to transcriptomic data at the level of tissue regions, where signals from multiple cells are combined. In the VOICE training pairs, each image and Xenium measurement correspond to an individual cell. The model therefore links a cell’s appearance to its molecular profile, while the library supplies information when morphology alone is insufficient.

To test whether the model works on new sections, the authors excluded five Xenium sections from model development. For each of these sections, they built the library of similar cells using other sections from the same tissue. In comparisons with three other methods, the authors report the best or joint-best performance across seven measures of single-cell gene activity prediction quality.

For a new H&E section, the model uses the appearance and location of its cells, while the library contributes gene activity estimates based on cells in which that activity has already been measured.