Xaira-Therapeutics/X-Atlas-Pisces

X-Atlas/Pisces

X-Atlas/Pisces is a Perturb-seq atlas containing 25.6 million perturbed single-cell transcriptomes across 16 biologically diverse contexts, including widely used cell lines (HCT116, HEK293T, HepG2), induced pluripotent stem cells (iPSCs), resting and CD3/CD28 activated Jurkat T lymphoma cells, and multi-lineage differentiating iPSCs.

Preprint: X-Cell: Scaling Causal Perturbation Prediction Across Diverse Cellular Contexts via Diffusion Language Models

Tutorial

from datasets import load_dataset

# load the entire dataset in streaming mode
ds = load_dataset("Xaira-Therapeutics/X-Atlas-Pisces", streaming=True)
# load only hct116
hct116_ds = load_dataset("Xaira-Therapeutics/X-Atlas-Pisces", streaming=True, split="HCT116")

Description of datasets

name	description (# cells)
HCT116	cells containing a valid dual-guide pair (3,409,169)
HEK293T	cells containing a valid dual-guide pair (4,534,299)
HepG2	cells containing a valid dual-guide pair AND perturbations in the 200-gene validation set (53,509) OR a non-targeting control (126,382)
iPSC	cells containing a valid dual-guide pair AND perturbations in the 200-gene validation set (83,580) OR a non-targeting control (233,426)
JurkatResting	cells containing a valid dual-guide pair (2,837,716)
JurkatActive	cells containing a valid dual-guide pair (2,795,555)

The 200-gene validation set is defined in Section 4.1.7 of the manuscript:

In iPSC and HepG2 screens from X-Atlas/Pisces, the 200 validation perturbations were randomly selected from gene targets meeting the following criteria across all
four genome-wide screens (HCT116, HEK293T, HepG2, and iPSC): (1) detectable expression in NTCs, (2) significant phenotype changes detected by the binary classifier
pipeline (FDR < 0.05), (3) at least 100 perturbed cells per screen, and (4) ≥ 50% knockdown efficiency.

Dataset metadata

The dataset contains the following information:

name	description
gene_token_id	gene identifiers corresponding to genes with non-zero expression in each cell. to be used with gene_expression. metadata/gene_metadata.parquet contains the mapping from gene_token_id to Ensembl ID and official gene symbol
gene_expression	raw counts for genes with non-zero expression. to be used with gene_token_id
cell_barcode	10X-generated cell barcode. the suffix -1 is replaced with -<SAMPLE>
sample	GEM batch
num_features	number of guides
num_guide_umis	number of UMIs for with each guide (guide a UMIs | guide b UMIs)
guide_target	guide identity
gene_target	gene targeted by guide
n_genes_by_counts	number of genes with non-zero counts
total_counts	total UMIs
total_counts_mt	total UMIs from MT genes
pct_counts_mt	% UMIs from MT genes
pass_guide_filter	boolean if cells contains two guides from the same guide pair

Gene metadata

All samples were aligned to the 10x Genomics GRCh38 2024-A pre-built reference genome (human reference (GRCh38) - 2024-A). Official gene symbols and ensembl IDs were extracted from the genes.gtf file.

# load metadata containing mappings to gene tokens and names
gene_metadata = load_dataset("Xaira-Therapeutics/X-Atlas-Pisces","gene_metadata")

name	description
ensembl_id	Ensembl ID
gene_name	official gene symbol
gene_token_id	gene identifiers corresponding to genes with non-zero expression in each cell. to be used with gene_token_id in the dataset

Additional resources

• X-Atlas/Orion dataset and tutorial: Hugging Face
• X-Atlas/Orion aggregated h5ads and additional metadata: Figshare

Citation

@article{wang2026xcell,
  title={X-Cell: Scaling Causal Perturbation Prediction Across Diverse Cellular Contexts via Diffusion Language Models},
  author={Wang, Chloe and Karimzadeh, Mehran and Ravindra, Neal G. and Bounds, Lexi R. and Alerasool, Nader and Huang, Ann C. and Ma, Shihao and Gulbranson, Daniel R. and Cui, Haotian and Lee, Yongju and Arjavalingam, Anusuya and MacKrell, Elliot J. and Wilken, Matthew S. and Chen, Jieming and Herken, Benjamin W. and Weber, Jesse A. and Onesto, Massimo M. and Gonzalez-Teran, Barbara and Leung, Nicole F. and Shi, Sally Yu and Smith, Byron J. and Lam, Sharon K. and Barner, Adam and Wright, Philip and Rumsey, Elizabeth M. and Kim, Soohong and Sit, Rene V. and Litterman, Adam J. and Chu, Ci and Wang, Bo},
  journal={bioRxiv},
  year={2026},
  url={https://www.biorxiv.org/content/10.64898/2026.03.18.712807v1}
}