Single-Cell Perturbation · arcinstitute
Perturb-Sapiens
A whole-organism atlas of predicted single-cell perturbation responses across 28 human tissues.
What Perturb-Sapiens is
Perturb-Sapiens is a database of predicted single-cell perturbation responses across the human body, developed by the Arc Institute. Designed as an evolving whole-organism atlas, it profiles cellular transcriptional changes caused by pharmacological and immunological perturbations across dozens of tissues.
Experimental single-cell perturbation screens are typically limited to select cell lines or peripheral blood mononuclear cells due to technical and biological constraints. Perturb-Sapiens addresses this limitation by using in-context foundation models to predict how perturbagens alter expression profiles systematically across diverse primary human cell types.
The dataset covers over 103 million cells across 28 human tissues and 40 cell classes. Researchers can use it to explore broad drug effects, evaluate cytokine signalling across distinct anatomical sites, and generate hypotheses for therapeutic target discovery.
How it was made
Perturb-Sapiens is generated using Stack-Large-Aligned, an in-context learning foundation model for single-cell biology. The model takes baseline expression profiles from Tabula Sapiens as query data and contextualises them using prompt data derived from Parse/OpenProblems peripheral blood mononuclear cell (PBMC) perturbation experiments.
The resulting data are organised by perturbation type into drug and cytokine directories stored in H5AD format. Drug perturbations include a dimethyl sulfoxide control, whilst cytokine perturbations include a phosphate-buffered saline (PBS) control. Cells in the dataset include generation logit scores (gen_logit) to assist in filtering high-quality predicted outputs.
Key facts
- Generating model
- Stack-Large-Aligned
- Total cells
- Over 103 million
- Human tissues
- 28
- Cell classes
- 40
- Drug conditions
- 111 (control: Dimethyl sulfoxide)
- Cytokine conditions
- 90 (control: PBS)
- Baseline query source
- Tabula Sapiens
- Prompt source
- Parse/OpenProblems PBMC perturbation data
- Data format
- H5AD
What you can use it for
- Drug response profiling across tissues
- Examine how 111 different drug treatments alter cellular gene expression systematically across multiple human tissue environments.
- Cytokine signalling analysis
- Investigate tissue-specific transcriptional responses to 90 distinct cytokine conditions to understand local immune signalling.
- Therapeutic target discovery
- Generate and prioritise mechanistic hypotheses for novel drug targets by analysing simulated perturbation outcomes before performing wet-lab screens.
Explore the data
No queryable preview
This dataset is not available in a queryable form — usually an unusual file format or a repository too large to convert. The original source may still be readable.Strengths and limits
Good at
- Provides a human whole-organism scope spanning 28 tissues and 40 cell classes.
- Supplies over 103 million predicted single-cell perturbation profiles.
- Delivers paired drug and cytokine conditions alongside matched experimental controls (DMSO and PBS).
- Integrates standard single-cell metadata, allowing straightforward ingestion into Scanpy workflows.
Watch out for
- Profiles are generated computationally by a foundation model and require experimental validation.
- May exhibit lower precision when predicting the effects of weak perturbations.
Questions
Can I look inside Perturb-Sapiens without downloading it?
Not here: it is not available in a queryable form. Its description and statistics are still shown.
What licence is Perturb-Sapiens under?
Its authors declare cc-by-4.0. Read the terms on the original page before reusing the data.
How were the perturbed cellular states generated?
They were generated by the Stack-Large-Aligned in-context foundation model, combining baseline cell transcriptomes from Tabula Sapiens with prompt data from Parse/OpenProblems PBMC perturbation screens.
What controls are available in the dataset?
Drug conditions use dimethyl sulfoxide (DMSO) as a baseline control, and cytokine conditions use phosphate-buffered saline (PBS).
What file format is used for the data?
The data are provided as H5AD files divided into 'drug' and 'cytokine' subdirectories, which can be read directly using standard single-cell libraries like Scanpy.
How should high-quality cells be filtered?
The dataset annotations include a generation logit field ('gen_logit'); filtering for values under 2.5 is recommended to retain high-quality predicted cells.
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 Perturb-Sapiens's own dataset card. Statements citing a figure that does not appear there are removed before publication. Licence and running details come from our registry.