Drug Differential Expression · tahoebio
tahoe-de-rhaister
Differential-expression summary statistics across 50 cancer cell lines and 384 drug perturbations.
What tahoe-de-rhaister is
tahoe-de-rhaister provides differential-expression summary statistics derived from the Tahoe-100M drug screen, produced by tahoebio. The dataset summarises observations from approximately 100 million single cells spanning 50 cancer cell lines treated with 384 compounds across three doses.
The resource addresses the need for pre-computed, gene-level perturbation responses without requiring researchers to process raw single-cell sequencing counts. It contains per-gene fold changes, statistical significance metrics, and pseudobulk expression deltas across cell line and treatment pairs, serving as a training and evaluation benchmark for the Rhaister model.
How it was made
The dataset aggregates single-cell expression profiles into two primary data structures: `pdex` and `cell_eval`. The `pdex` table stores long-format differential expression results per cell line, treatment, gene, and plate relative to plate-matched DMSO controls, using Mann–Whitney U tests and Benjamini–Hochberg false discovery rate corrections. The `cell_eval` table provides wide-format pseudobulk expression deltas on a linear normalised scale across an analysis set of approximately 2,000 genes.
The repository also provides control-state reference tables including DMSO-control mean expression, whole-cell centroid embeddings, and highly variable gene centroids. Dataset configuration files define plate aggregation rules, the 2,000-gene analysis set, and compositional holdout splits where individual cell lines and treatments appear in training but specific cell-treatment combinations are reserved for evaluation.
Key facts
- Publisher
- tahoebio
- Cell lines
- 50 cancer cell lines (Cellosaurus IDs)
- Compounds
- 384 drugs across 3 concentrations
- Underlying cells
- ~100 million single cells
- Row count (pdex)
- ~4.1 billion rows
- Primary modalities
- Tabular, text
- Primary tables
- pdex, cell_eval, control_expression, cell_centroids, cell_centroids_hvg
- Gene feature space
- ~2,000 genes in pseudobulk evaluation sets
What you can use it for
- Machine learning model training
- Train perturbation prediction models such as Rhaister using pre-computed gene-level differential statistics and pseudobulk deltas.
- Zero-shot perturbation evaluation
- Benchmark zero-shot cellular response predictions against baseline control tables and predefined compositional holdout splits.
- Drug screening analysis
- Query specific compound-by-concentration impacts across 50 cancer cell lines using Mann–Whitney U test statistics, fold changes, and FDR values.
Explore the data
The schema, real rows, full-text search and a SQL filter — read in your browser, nothing downloaded.
Explore
Read the real rows without downloading anything
Reading rows…
LiveRead from tahoebio/tahoe-de-rhaister at the moment you asked. Nothing is cached or stored — every row above came from that request.
Strengths and limits
Good at
- Summarises high-volume single-cell perturbation screens into accessible pseudobulk deltas and statistical tables.
- Includes plate-matched DMSO control comparisons with adjusted p-values and cell count memberships.
- Provides standardised compositional holdout splits designed for evaluating combinatorial generalisation.
Watch out for
- Evaluations in the wide format are restricted to a defined subset of approximately 2,000 genes.
- Log2 fold changes produce negative infinity values when target expression is zero.
- Data structure relies on streaming or chunked loading due to the large scale of roughly 4.1 billion rows in the long-format table.
Questions
Can I look inside tahoe-de-rhaister without downloading it?
Yes. This page reads its schema, real rows and per-column statistics in the browser, with full-text search and a SQL filter. Nothing is downloaded to your machine.
What licence is tahoe-de-rhaister under?
Its authors declare cc0-1.0. Read the terms on the original page before reusing the data.
What is the difference between the pdex and cell_eval tables?
The pdex table is a long-format dataset containing statistical summaries such as log2 fold change, percent change, Mann–Whitney U p-values, and FDR for each gene and plate. The cell_eval table is a wide-format dataset giving the pseudobulk delta (mean treated minus mean control) across approximately 2,000 genes per treatment condition.
How are controls matched for differential expression calculations?
Differential expression metrics in the pdex table are computed relative to plate-matched DMSO control wells from the corresponding 96-well plate.
What splits are provided for testing model generalisation?
The repository provides compositional holdout split configurations (such as 5_holdout through 9_holdout), where specific cell line and treatment pairs are held out together while ensuring each cell line and treatment is observed individually during training.
How to cite it
@article{zhang2025tahoe, title={Tahoe-100M: A Giga-Scale Single-Cell Perturbation Atlas for Context-Dependent Gene Function and Cellular Modeling}, author={Zhang, Jesse and Ubas, Airol A and de Borja, Richard and Svensson, Valentine and Thomas, Nicole and Thakar, Neha and Lai, Ian and Winters, Aidan and Khan, Umair and Jones, Matthew G and others}, journal={bioRxiv}, pages={2025--02}, year={2025}, publisher={Cold Spring Harbor Laboratory}}
Written from tahoe-de-rhaister'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.