Multi-Modal Foundation · arcinstitute
Stack-Large Aligned
A 217M-parameter single-cell foundation model aligned for cross-batch consistency and perturbation analysis.
What Stack-Large Aligned is
Stack-Large Aligned is a 217-million-parameter encoder-decoder foundation model for single-cell biology developed by Arc Institute. It is built to model cellular transcriptomes using cell-by-gene matrix chunks as its fundamental input unit.
This model variant is specifically fine-tuned to maintain cross-batch consistency across datasets. Through in-context learning, it supports multi-study perturbation analysis and downstream integration tasks, enabling researchers to generalise biological responses and generate unseen cellular profiles in novel experimental contexts.
How it works
The model is built on a tabular attention architecture that alternates between cell-wise and gene-wise attention mechanisms. This design permits information to circulate both internally within individual cells and externally across groups of cells. By processing cell-by-gene matrix chunks, the network leverages in-context learning to model context-dependent cellular behaviours.
Pretraining was conducted on the human scBaseCount dataset comprising approximately 150 million cells over 10 epochs. The model was then aligned over 8 epochs on a combined dataset consisting of 45 million cells from CellxGene and 10 million peripheral blood mononuclear cells (PBMC) from Parse.
Key facts
- Architecture
- Tabular Attention encoder-decoder (alternating cell-wise and gene-wise attention)
- Parameters
- 217M
- Input unit
- Cell-by-gene matrix chunks
- Pretraining data
- Full human scBaseCount (~150M cells)
- Pretraining duration
- 10 epochs
- Alignment data
- CellxGene 45M + Parse 10M PBMC
- Alignment duration
- 8 epochs
What you can use it for
- Cross-Batch Data Alignment
- Aligning heterogeneous single-cell transcriptomic datasets to ensure consistency across experimental batches.
- Multi-Study Perturbation Analysis
- Evaluating the biological impacts of experimental perturbations across separate single-cell studies.
- In-Context Profile Generation
- Predicting and generating unseen cellular expression profiles within novel biological contexts using in-context examples.
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
- Captures both intra-cellular and inter-cellular information flow using alternating cell-wise and gene-wise attention.
- Supports in-context learning to generalise biological effects to previously unobserved cellular contexts.
- Fine-tuned specifically for cross-batch consistency across multiple large-scale single-cell repositories.
Watch out for
- Pretraining is limited to human single-cell transcriptomic data from the scBaseCount dataset.
- Alignment fine-tuning is restricted to data derived from CellxGene and Parse PBMC collections.
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 Aligned 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 feature of Stack?
Stack employs a tabular attention architecture that alternates between cell-wise and gene-wise attention, enabling the model to learn relationships both within and between cells.
How does Stack-Large Aligned differ from other Stack variants?
Stack-Large Aligned is specifically fine-tuned for 8 epochs on 45 million CellxGene cells and 10 million Parse PBMC cells to optimise cross-batch consistency for downstream integration and multi-study analysis.
What input data format does the model require?
The model processes data structured into cell-by-gene matrix chunks as its foundational input unit.
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 Aligned'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.