Multi-Modal Foundation · arcinstitute
STACK Large
A single-cell transcriptomics model for in-context learning across cell types and biological perturbations.
What STACK Large is
STACK Large is a 217-million-parameter encoder-decoder foundation model for single-cell biology developed by the Arc Institute. Designed to process single-cell transcriptomics data, the model sets cell-by-gene matrix chunks as its fundamental input unit.
The model addresses the challenge of predicting cellular behavior across diverse states and interventions. By supporting in-context learning, it allows computational biologists and genomics researchers to generalise biological effects across distinct cell types and generate expression profiles for unseen cells in novel contexts.
How it works
STACK Large is built on a tabular attention architecture that alternates between cell-wise and gene-wise attention. This structure facilitates both intracellular (gene-to-gene) and intercellular (cell-to-cell) information flow across matrix chunks of expression measurements.
The model was pre-trained for 10 epochs on the full human scBaseCount dataset, which encompasses approximately 150 million cells.
Key facts
- Architecture
- Tabular Attention encoder-decoder (alternating cell-wise and gene-wise attention)
- Parameter count
- 217M
- Pre-training data
- Full human scBaseCount (~150M cells)
- Pre-training epochs
- 10
What you can use it for
- Single-cell representation learning
- Extracting latent biological embeddings from single-cell transcriptomic profiles.
- Perturbation response prediction
- Generalising biological responses and perturbation states across various cell types using in-context examples.
- Novel context profile generation
- Generating unobserved single-cell expression profiles within unseen biological contexts.
Run it here
We list this model but cannot run it yet: no pipeline_tag — needs a human.
Everything on this page still applies. Running it needs a loader for this repository, not a bigger machine.
Strengths and limits
Good at
- Alternates cell-wise and gene-wise attention to capture both intra- and inter-cellular dynamics.
- Pre-trained on a broad corpus of approximately 150 million human cells across 10 epochs.
- Supports in-context learning directly across cell types and perturbation conditions.
Watch out for
- Pre-training data is restricted to human single-cell records from the scBaseCount collection.
- Expects structured cell-by-gene matrix chunks rather than standard isolated cell vectors.
Inputs and outputs
You provide
Sequence or text
A protein or nucleotide sequence, or free text — whichever this model was trained on.
You get back
Mean-pooled representation, one vector per input.
Licence
Check the licence before commercial use · See model card
The See model card licence carries terms worth reading before you depend on it.
Questions
Can I use STACK Large commercially?
Check the licence before commercial use. The See model card licence carries terms worth reading before you depend on it.
What is the primary architectural mechanism behind STACK Large?
It relies on tabular attention, an encoder-decoder framework that alternates between gene-wise and cell-wise attention across chunks of cell-by-gene matrices.
What dataset was used during pre-training?
The model was pre-trained for 10 epochs on the human scBaseCount dataset, comprising roughly 150 million cells.
How does the model handle unseen biological contexts?
It utilizes in-context learning over cell-by-gene inputs to generalize biological effects and generate cell profiles in previously unobserved conditions.
How to cite it
Dong et al., 2026: Stack: In-context modeling of single-cell biology. Preprint. https://www.biorxiv.org/content/10.64898/2026.01.09.698608v1
Written from STACK Large's own documentation (model card). Statements citing a figure that does not appear there are removed before publication. Licence and running details come from our registry.