tahoebio/tahoe-de-rhaister

Tahoe-100M Drug Screen Differential Expression Summary Statistics

Differential-expression summary statistics for the Tahoe-100M drug screen including 50 cancer cell lines × 384 drugs × 3 doses, one of the training datasets for Rhaister. These are the statistics computed from the ~100M raw single cells of tahoebio/Tahoe-100M, including per-gene fold changes, significance, and pseudobulk deltas per (cell line, treatment).

The repo includes two tables (pdex and cell_eval) plus the control-state references used for zero-shot evaluation, and the dataset-definition files (filtering / split criteria) under definition/.

pdex per-gene differential expression (long format)

from datasets import load_dataset
pdex = load_dataset("tahoebio/tahoe-de-rhaister", "pdex", streaming=True, split="train")

One row per (cell line, treatment, gene, plate). ~4.1 B rows total.

Column	Description
cell_line	Cellosaurus ID of the cancer cell line.
target	Treatment: compound × concentration, e.g. [('5-Azacytidine', 0.05, 'uM')].
plate	96-well plate ID (1–14); the unit of plate-matched DMSO control.
feature	Gene symbol.
fold_change	log2 fold change, log2(target_mean / ref_mean). -inf where target_mean = 0.
percent_change	(target_mean − ref_mean) / ref_mean.
p_value	Mann–Whitney U p-value, treated vs. plate-matched DMSO.
statistic	Mann–Whitney U statistic.
fdr	Benjamini–Hochberg adjusted p-value.
target_mean / ref_mean	Mean expression in the treated / control group.
target_membership / ref_membership	Cell counts in the treated / control group.

cell_eval pseudobulk expression deltas (wide format)

cell_eval = load_dataset("tahoebio/tahoe-de-rhaister", "cell_eval", streaming=True, split="train")

One row per (cell line, treatment); 14 per-plate files. Columns cell_line, treatment, then one column per gene (~2,000) holding the pseudobulk delta mean(treated) − mean(control) on the linear (normalized) scale.

Zero-shot reference tables

Control-state references used as the test-time predictor in the zero-shot setting:

• control_expression: per (cell line, plate, gene) DMSO-control mean expression.
• cell_centroids: one DMSO-condition centroid embedding per cell line.
• cell_centroids_hvg: per (cell line, plate) highly-variable-gene centroid of DMSO-condition mean expression.

Definition & filtering criteria

definition/ contains the dataset definition (dataset.toml) and the holdout split TOMLs (5_holdout–9_holdout and titration variants). These encode the data paths, the 2,000-gene analysis set, plate aggregation (mean fold change / ref_mean, min p-value / FDR), and the compositional tahoe_5_holdout split. A (cell, treatment) pair is held out only when its cell line is in the holdout set and its treatment is in that cell's held-out list, so each test cell and treatment is still seen individually in training.

Source

@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}
}