What can you ask with your data?
Pick the kind of data you have. Each one opens a different set of questions, all answered by one model.
H&E slide only
A routine pathology slide, no sequencing.
- Find tumor regions
- Label tissue types across the slide
- Predict which genes each region expresses
Spatial transcriptomics
H&E plus measured genes for every cell.
- Assign a cell type to every cell
- Use genes alone, or genes with morphology
- See where each cell type sits in the tissue
Single-cell genes
A ranked gene list for each cell.
- Name the cell type from its top genes
- Fill in genes that were not measured
H&E slide only
Two TCGA lung cancer slides with no sequencing and no annotation. The model looked at every 128 µm patch and answered three questions: is this tumor, what tissue is this, and which genes would it express.
Spatial transcriptomics
A public 10x Xenium lung cancer section, where each cell has measured genes and an H&E image of the same tissue. The model reads each cell's top genes, optionally together with its H&E crop, and names the cell type.
Single cell
Tumor samples from seven lung cancer patients (GSE127465). Each cell was reduced to its 20 most expressed genes, in order, and the model was asked for the cell type. The authors published their own labels, so you can compare the two directly.
Each dot is a cell, placed where the authors' own map put it. Where colors match across the handle, the model agreed with the published label.
Ask the model
Pick a question. These are real answers the model gave, saved ahead of time so the page runs without a GPU.
How it compares
The paper gave the same inputs and instructions to ten other medical and general-purpose AI models, including the model VGL-stGen was built from. Pick a question type.
Good to know
- Trained and tested on lung tissue only. Other tissues need their own validation.
- For research use. Tumor regions are model scores, not a diagnosis.
- Nothing runs live. All results were produced with the released model and the scripts in this repository.
Shin H., Lee D., Bae S., Na K. J., Lee D., Choi H. VGL-stGen: Learning a shared multimodal model for lung histology and spatial transcriptomics. arXiv:2610.28266, 2026. Data from TCGA, 10x Genomics, and GEO GSE127465.