altoslabs/primeflow-vcc-datasets

PRiMeFlow VCC Datasets

This repository contains the preprocessed datasets, data splits, and gene features used in:

PRiMeFlow: Capturing Complex Expression Heterogeneity in Perturbation Response Modelling Zichao Yan, Yan Wu, Mica Xu Ji, Chaitra Agrahar, Esther Wershof, Marcel Nassar, Mehrshad Sadria, Ridvan Eksi, et al. arXiv:2604.13986

PRiMeFlow is an end-to-end flow-matching model of genetic and small-molecule perturbation responses that operates directly in gene expression space. The data here supports the pretraining–finetuning strategy described in the paper, and evaluation on the H1 human embryonic stem cell data from the ARC Virtual Cell Challenge (VCC). Code to reproduce the results is available in the PRiMeFlow GitHub repository.

Contents

Preprocessed expression data (.h5ad.gz)

All expression datasets are provided as gzipped AnnData files. Cell identifiers are stored in .obs_names and are referenced by the split files below.

File(s)	Description	Size
cd4t_perturbseq_D{1-4}_{Rest,Stim8hr,Stim48hr}_preprocessed.h5ad.gz	Genome-scale Perturb-seq in primary human CD4T cells across four donors (D1–D4), each at rest and at 8 hr / 48 hr post-stimulation (12 files; Zhu, Dann et al., 2025)	~198 GB
vcc_train_preprocessed.h5ad.gz	ARC Virtual Cell Challenge H1 hESC training data	4.3 GB
replogle22_k562_preprocessed.h5ad.gz	Genome-scale Perturb-seq, K562 (Replogle et al., 2022)	13.7 GB
replogle22_rpe1_preprocessed.h5ad.gz	Genome-scale Perturb-seq, RPE1 (Replogle et al., 2022)	1.9 GB
replogle20_preprocessed.h5ad.gz	Direct-capture Perturb-seq (Replogle et al., 2020)	253 MB
nadig24_hepg2_preprocessed.h5ad.gz	HepG2 essential-gene Perturb-seq (Nadig et al., 2025)	1.3 GB
nadig24_jurkat_preprocessed.h5ad.gz	Jurkat essential-gene Perturb-seq (Nadig et al., 2025)	1.9 GB
jiang24_preprocessed.h5ad.gz	Molecular pathway signature reconstruction screens (Jiang et al., 2024)	10 GB
mcfaline23_gxe_preprocessed.h5ad.gz	Chemical–genetic (GxE) single-cell screens (McFaline-Figueroa et al., 2024)	3.2 GB
feng24_preprocessed.h5ad.gz	Genome-scale single-cell CRISPRi map of trans gene regulation across human iPSC lines (Feng et al., 2026)	10.3 GB

Note on disk space. The .h5ad.gz files total ~244 GB compressed. Most of this comes from the twelve CD4T cell datasets (~198 GB). Plan storage accordingly. Refer to our download script in the PRiMeFlow codebase.

Gene features

• ESM2_pert_features.parquet (595 MB): ESM2 protein language model embeddings used to represent perturbed genes. These are derived from the Arc Institute's competition support set. A pre-processed copy is provided here for convenience.
• vcc_filtered_genes_18001.csv: the filtered set of 18,001 genes used as the model's expression space for the VCC task.

Data splits

Splits are provided as .csv files with two columns: the cell ID (matching .obs_names of the corresponding .h5ad files) and the split assignment.

• vcc_split_multifile_pretrain.csv: pretraining split spanning the full multi-dataset atlas.
• vcc_split_multifile_pretrain_nocd4t.csv: pretraining split excluding the CD4T cell datasets.

Prediction dataframes

• prediction_dataframe_vcc_test_multifile_h5.csv: perturbation-covariate specification for generating VCC test predictions.
• prediction_dataframe_vcc_controls_only_multifile_h5.csv: control-cell only variant of the above.

Legacy files

• pretrain_legacy/ — earlier versions of the pretraining files, kept for reproducibility of intermediate results. New users should ignore this folder.

Source publications

The preprocessed datasets are derived from the following publications. If you use them, please cite the original works alongside the PRiMeFlow paper:

• Replogle, J. M., et al. (2020). Combinatorial single-cell CRISPR screens by direct guide RNA capture and targeted sequencing. Nature Biotechnology, 38(8):954–961.
• Replogle, J. M., et al. (2022). Mapping information-rich genotype–phenotype landscapes with genome-scale Perturb-seq. Cell, 185(14):2559–2575.
• Jiang, L., et al. (2024). Systematic reconstruction of molecular pathway signatures using scalable single-cell perturbation screens. Nature Cell Biology, 27(3):505-517.
• McFaline-Figueroa, J. L., et al. (2024). Multiplex single-cell chemical genomics reveals the kinase dependence of the response to targeted therapy. Cell Genomics, 4(2).
• Nadig, A., et al. (2025). Transcriptome-wide analysis of differential expression in perturbation atlases. Nature Genetics, 57(5):1228–1237.
• Feng, C., et al. (2026). A genome-scale single-cell crispri map of trans gene regulation across human pluripotent stem cell lines. Cell Genomics, 6(2).
• Zhu, R., et al. (2025). Genome-scale perturb-seq in primary human CD4+ T cells maps context-specific regulators of T cell programs and human immune traits. bioRxiv, 2025.12.23.696273.
• Roohani, Y. H., et al. (2025). Virtual Cell Challenge: Toward a Turing test for the virtual cell. Cell, 188(13):3370–3374.

Citation

@article{yan2026primeflow,
  title   = {PRiMeFlow: Capturing Complex Expression Heterogeneity in Perturbation Response Modelling},
  author  = {Yan, Zichao and Wu, Yan and Ji, Mica Xu and Agrahar, Chaitra and Wershof, Esther and Nassar, Marcel and Sadria, Mehrshad and Eksi, Ridvan and Trifonov, Vladimir and Ibarra, Ignacio and others},
  journal = {arXiv preprint arXiv:2604.13986},
  year    = {2026}
}

License

The curation, preprocessing, data splits, and packaging in this repository (i.e., the compilation itself) are released under the Creative Commons Attribution 4.0 International (CC BY 4.0) license. This license applies to the compilation only and does not override or diminish the terms attached to any underlying dataset.

When you use this data, please (a) cite the relevant source publications alongside PRiMeFlow, and (b) indicate that the data has been modified — it has been re-preprocessed from the original deposits.

Third-party licenses

Two of the bundled datasets were released under the MIT License by their original authors and remain under that license. Their upstream notices are preserved in the LICENSES/ directory:

• feng24_preprocessed.h5ad.gz — Feng et al. (2026), figshare → LICENSES/feng24-MIT.txt
• cd4t_perturbseq_* — Zhu et al. (2025), CZI CELLxGENE / virtualcellmodels → LICENSES/cd4t_perturbseq-MIT.txt