CRISPR Differential Expression · tahoebio
replogle-nadig-de-rhaister
Differential expression statistics across four cell lines and ~2,000 CRISPR knockdowns used to train Rhaister.
What replogle-nadig-de-rhaister is
replogle-nadig-de-rhaister is a tabular genomics dataset published by tahoebio containing differential-expression summary statistics from the Replogle-Nadig genome-scale CRISPR screen. It covers approximately 2,000 gene knockdowns evaluated across four human cell lines: HepG2, Jurkat, K562, and RPE1. The dataset was prepared to train Rhaister, a computational model designed to predict perturbation outcomes.
The dataset translates raw single-cell RNA sequencing measurements into structured differential-expression metrics and pseudobulk profiles. By aggregating single-cell observations into fold changes, non-parametric test statistics, and expression shifts relative to non-targeting controls, it enables downstream modeling of genetic interventions without requiring raw read-level processing.
How it was made
The data are derived from single-cell CRISPR knockdown screens stratified by cell line and sequencing batch (gem groups 1 to 56). For each target gene knockdown, expression values are compared against non-targeting reference controls to compute per-gene metrics, including log2 fold changes, percentage changes, mean expressions, and cell counts. Statistical significance is evaluated using the Mann–Whitney U test alongside Benjamini–Hochberg false discovery rate (FDR) adjustments.
The repository provides two primary representations. The first is 'pdex', a long-format table containing roughly 1.5 billion rows indexed by cell line, gem group, target gene, and measured feature. The second is 'cell_eval', a wide-format pseudobulk table reporting the linear delta (perturbed mean minus control mean) across approximately 6,546 measured genes per condition. Dataset configuration files define a compositional train/test split across four holdout splits (split_0 to split_3) alongside a subset of 2,000 highly-variable genes.
Key facts
- Publisher
- tahoebio
- Cell lines
- HepG2, Jurkat, K562, RPE1
- Perturbations
- ~2,000 CRISPR gene knockdowns
- pdex row count
- ~1.5 billion rows
- cell_eval feature count
- ~6,546 measured genes
- Sequencing batches
- gem_group 1–56
- Modality
- Tabular summary statistics
- Model application
- Training data for Rhaister
What you can use it for
- Perturbation model training
- Train predictive machine learning architectures such as Rhaister using precomputed pseudobulk shifts and differential statistics across diverse knockdowns.
- Comparative response analysis
- Compare gene expression changes, fold changes, and significance statistics induced by identical target knockdowns across the four distinct cell lines.
- Compositional generalization benchmarking
- Evaluate model generalisation to unseen cell line and target gene combinations using the predefined compositional splits.
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/replogle-nadig-de-rhaister at the moment you asked. Nothing is cached or stored — every row above came from that request.
Strengths and limits
Good at
- Provides precomputed statistical tests and FDR values, eliminating the computational burden of processing single cells from scratch.
- Features both granular long-format per-gene tests (~1.5 billion rows) and wide-format linear pseudobulk deltas.
- Includes predefined compositional holdout splits designed so that each test cell line and target are observed individually during training.
Watch out for
- Coverage is restricted to four specific cell lines (HepG2, Jurkat, K562, and RPE1).
- Perturbations are limited to approximately 2,000 targeted gene knockdowns rather than full genome-wide targets.
- Fold change values evaluate to negative infinity when target expression drops to zero.
Questions
Can I look inside replogle-nadig-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 replogle-nadig-de-rhaister under?
Its authors declare cc-by-4.0. Read the terms on the original page before reusing the data.
What is the difference between the pdex and cell_eval configurations?
The pdex configuration provides a long-format table of around 1.5 billion rows containing per-gene differential statistics such as p-values, FDR, and fold changes. The cell_eval configuration provides a wide-format pseudobulk table reporting linear expression deltas across approximately 6,546 measured genes.
How are statistical p-values calculated in this dataset?
Significance is determined using a two-sided Mann–Whitney U test comparing perturbed cells against non-targeting controls, with Benjamini–Hochberg adjustment applied to calculate FDR values.
What cell lines are included in the screens?
The dataset contains data from four human cell lines: HepG2, Jurkat, K562, and RPE1.
How are the evaluation splits constructed?
The dataset includes four compositional holdout splits (split_0 to split_3) where a (cell line, gene) pair is held out only if the cell line is in the holdout set and the target is on that line's holdout list, ensuring both targets and lines are seen during training.
Written from replogle-nadig-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.