tx_evaluation

Evaluation datasets for the tx_evaluation benchmark: how well does a transcriptomics embedding recover known biology?

Three genome-scale CRISPRi Perturb-seq screens over the same 2,393 target genes, in three different human cell lines.

adata.obsm is empty by design. These files carry raw counts and metadata only. You run your model over the cells, inject the resulting vectors, and the benchmark scores them.

Files

File	Cell line	Cell type	Cells	Genes	Perturbations	Batches	Source
replogle_2022_rpe1_e.h5ad	RPE1	epithelial	247,914	8,749	2,393	56	Replogle et al. 2022
nadig_2024_hepg2.h5ad	HEPG2	hepatoblastoma	145,473	9,624	2,393	56	Nadig et al. 2024
nadig_2024_jurkat.h5ad	Jurkat	T lymphoblast	262,956	8,882	2,393	55	Nadig et al. 2024

~3.6 GB total. All human, all CRISPRi, all essential-gene screens. Because the perturbation set is shared, scores are directly comparable across the three cell lines.

Contents of each file

• X — raw integer UMI counts, highly-variable-gene filtered, stored as float32 (dense).

• obs — per-cell metadata:

  column	meaning
  gene_name	perturbation label: the CRISPRi target gene, or non-targeting for controls
  is_control	boolean; controls are 3–5% of cells
  dataset_batch_num	experimental batch — the batch-effect axis the benchmark corrects for
  cell_line, cell_type, organism, disease, perturbation_type	constant per file
  sgID_AB, transcript, ensembl_gene_id	guide and target identifiers
  UMI_count, gene_count, percent_mito, percent_ribo, mitopercent	per-cell QC

• var — per-gene metadata: gene_name, ensembl_gene_id, chr, start, end, strand, length, type_of_gene, and per-gene mean / std / cv / fano.

• obsm, layers, uns, varm, obsp, varp — all empty.

Usage

from huggingface_hub import hf_hub_download
import anndata

path = hf_hub_download(
    repo_id="bendidiihab/tx_evaluation",
    filename="nadig_2024_hepg2.h5ad",
    repo_type="dataset",
)
adata = anndata.read_h5ad(path)

print(adata)                                  # 145473 x 9624
print(adata.obs.gene_name.nunique())          # 2394 (2393 targets + non-targeting)
print(len(adata.obsm))                        # 0 -- bring your own

Full pipeline:

git clone https://github.com/IhabBendidi/tx_evaluation.git && cd tx_evaluation
pip install -e .

python scripts/download_datasets.py           # all three files into datasets/eval/

python -m biomodalities.data.inject_embeddings \
    --h5ad datasets/eval/nadig_2024_hepg2.h5ad \
    --npy  my_model_hepg2.npy --key my_embedding --inplace

python -m biomodalities.data.data_split       # train / test1 / test2
python -m biomodalities.data.update_configs --keys my_embedding
bash cfg/schedulers/run_all_jobs.sh

Your .npy must be (n_obs, n_dims) and in the same cell order as the h5ad.

What the benchmark measures

Metric	Question
Gene Relationship Recall	Do functionally related genes embed close together? (CORUM, HuMAP, SIGNOR, Reactome, StringDB)
Perturbation Consistency	Are replicates of a perturbation more self-similar than resampled controls?
KNN Separability	Can a k-NN recover the perturbation label on held-out batches?
Linear Separability	Can a linear probe?
Linear Invertibility	Can expression be decoded back, for perturbations never seen in training?
Batch iLISI	How much batch signal remains? (lower is better)

Provenance and preprocessing

These are re-releases of published data, harmonised to a common schema: consistent obs column names across all three files, a shared perturbation vocabulary, and HVG filtering per dataset. Counts are otherwise unmodified — no normalisation, no log transform. Normalisation happens inside the benchmark where a task calls for it.

Licensing

Redistributed under CC-BY-4.0, consistent with the terms of the source publications. Please cite the original work:

@article{replogle2022mapping,
  title   = {Mapping information-rich genotype-phenotype landscapes with genome-scale Perturb-seq},
  author  = {Replogle, Joseph M and Saunders, Reuben A and Pogson, Angela N and others},
  journal = {Cell},
  volume  = {185},
  number  = {14},
  pages   = {2559--2575},
  year    = {2022}
}

@article{nadig2024transcriptome,
  title   = {Transcriptome-wide characterization of genetic perturbations},
  author  = {Nadig, Ajay and Replogle, Joseph M and Pogson, Angela N and others},
  year    = {2024}
}

Links

• Code: https://github.com/IhabBendidi/tx_evaluation