↗ EXPAND
FROM MORPHOLOGY TO MOLECULES
Every cell has
a story.
Give it a voice.
Reveal the molecular landscape within routine H&E. A multimodal foundation model for single-cell transcriptomic reconstruction.
University of Michigan · Weill Cornell Medicine
HISTOLOGY → MOLECULAR INSIGHTVOICE / 25M
A closer look at the cellular landscape H&E
25.4M
cells used for training
18,580
output genes
15
training tissue types
3.58M
held-out evaluation cells
01 / SELECTED APPLICATIONS
See biology in context.
From the architecture of an islet to the molecular programs of disease. Selected results from the VOICE manuscript.
02 / THE MODEL
One image.
Multiple layers of insight.
VOICE combines morphology, spatial context and measured reference profiles. Its predictions are molecular scores—not measured molecule counts.
STAGE 01
→Align
Cell-resolved histology representations align with scGPT transcriptomic features through contrastive learning.
STAGE 02
→Contextualize
A spatial-attention decoder models cells and their neighbors to predict expression and generate cell embeddings.
STAGE 03
Retrieve & fuse
Gene-specific weights combine direct predictions with expression retrieved from reference-cell clusters.
03 / VOICE WORKBENCH
Your tissue.
A new perspective.
Upload a small H&E region, enter its physical resolution, and reconstruct cell-level expression with cloud GPU inference.
Checking inference service…PNG / JPEG / TIFF · up to 1024 × 1024 px · 20 MB
01
Cell boundaries
Predicted nuclei and source-image coordinates.
02
Expression maps
Direct and reference-fused gene scores in AnnData.
03
Cell embeddings
Stage 2 image features and spatial features as NPZ.
Research use only. Use de-identified images. Processing runs on Hugging Face; free GPU access is subject to availability and daily quotas. Whole-slide analysis is available through the command-line pipeline.
OPEN SCIENCE / VOICE
Histology-Based Reconstruction of
Single-Cell Transcriptomics through
multimodal foundation models
Xin Luo*, Haoxuan Zeng* and colleagues