01 / IMAGE COLLECTION
Image Atlas
231M pretraining cell images across 40 species.
Explore real images Full Image Atlas · available
A foundation model for single-cell biology
CellNet: A panoramic view
of the cell
DNA-informed images. Panoramic transcriptome modelling. CellNet brings gene identity and RNA expression together to connect cellular diversity with molecular detail.
02 / PANORAMIC LEARNING
Foundation Model
DNA-informed coding and non-coding genes, in one panoramic cell context.
Model & weights
Built for
biology at scale.Pretraining collection
231Mcell profiles
40species
1.01Mgene identities
01THE IDEA
From molecules to a cellular panoramaGive every gene
a place in the picture.
DNA-derived identities extend coverage across coding and non-coding genes. Image patches bring many genes into each token, making panoramic transcriptome modelling computationally practical.
01
Start with identity
DNA-derived representations describe genes, including those without a protein product, and define a species-specific spatial layout.
02
Make expression visible
Measured RNA expression fills a 384 × 384 canvas. A separate measurement mask preserves the difference between an observed zero and an unmeasured gene.
03
Learn the cellular context
Multiview latent learning builds cell representations. A gene-conditioned decoder connects that context back to individual gene expression.
A closer look at the architecture
02THE IMAGE ATLAS
24 real cells · 3 speciesReal cells.
A different kind of image.
Each image is a real single-cell transcriptome, mapped onto its native gene layout. Explore the patterns, then switch to the measurement mask to see which genes were observed.
Loading real transcriptome images…
EXPRESSIONLow → high
Expression uses the magma colormap with display gamma 0.25; the mask shows measured genes in gold. Select any cell to inspect both views and the original linear rendering.
03THE RESEARCH
Evaluation protocols ↗From cellular identity
to molecular detail.
One framework for studying which cells are present, how they differ, and what their genes reveal. Explore the applications evaluated in the CellNet manuscript.
REPRESENT CELLS
See the rare. Resolve the state.
CellNet supports cell annotation, reference mapping and integration, with evaluations of rare populations and condition-associated states.
0.646Kidney · rare-class Macro-F1
0.762Lung · rare-class Macro-F1
Donor-disjoint evaluation, reported separately for the kidney and lung cohorts.
Explore Figure 2 ↗
RECOVER EXPRESSION
Explore beyond measured genes.
Expression prediction extends biological exploration beyond the genes captured by an assay. From coding genes to selected non-coding transcripts, CellNet enables the study of unmeasured expression in cell states and spatial patterns.
Explore Figure 3 ↗
EXPLORE GENE RELATIONSHIPS
Non-coding RNA. Functional connections.
Use non-coding RNA as a starting point to retrieve associated gene programmes. Connect underexplored transcripts to developmental and neural biology through cellular context.
Try the association example ↗04OPEN RESOURCES
Made to be explored. Built to be reused.Your next discovery
starts here.
Start with the code, explore real examples, or download the full Image Atlas.
CellNet, open source.
Model weights, fitted layouts, inference tools and runnable task examples in one repository.
Python ≥ 3.10 · Git LFS · Apache-2.0 codeOpen GitHub ↗Small data. Real cells.
24 real cells from three species, with expression arrays, measurement masks, gene identities and a file catalog.
24 cells · 3 species · ZIP · ≈ 8.74 MiBDownload the example ↓Image Atlas
Download the full Image Atlas, measurement masks and fitted gene-to-image maps, with a step-by-step setup guide.
231M cell images · 40 speciesDownload Image Atlas · code: xabe ↗Explore the release files
GET STARTED
Start panoramic modelling.
Install Git LFS first. The repository includes weights and fitted assets; the Python wheel contains code.
Read the quick start ↗Terminal
# Clone code, weights and fitted layouts
git lfs install
git clone https://github.com/daydayup-xq/CellNet.git
cd CellNet
git lfs pull
pip install -e '.[h5ad]'THE WORK BEHIND THE MODEL
Panoramic modelling of single-cell transcriptomes at scale with CellNet
* Equal contribution · arXiv:2610.21943
Hunan UniversityUniversity of Electronic Science and Technology of ChinaXiamen UniversityYuelushan Laboratory