PertResolve-Bench datasets

This repository contains the allele-resolved data and the public perturbation datasets used by the PertResolve manuscript. It is a data release only: it does not contain the manuscript figures, analysis code, model predictions, or the results/ directory. The corresponding code and result reports remain in the PertResolve GitHub repository.

Included data

Allele-resolved benchmark

• benchmark/allele_perturb_bench.csv: 470 protein-coding variant conditions plus two wild-type reference rows, with gene, protein, biophysical features, cell counts and evaluation partitions.
• benchmark/hotspot_external_definition.txt: external hotspot definition used by the benchmark annotations.
• benchmark/supplementary_data_1_datasets.xlsx: the Supplementary Data 1 inventory and source metadata.

The complete compact allele package is included:

• TP53 and KRAS (GSE161824): processed cell-by-gene matrices, cell and gene tables, cell metadata and variants2cell assignments under allele/ursu_gse161824/. Source: GEO GSE161824.
• GATA1 (GSE215253/PerturbNet processing): allele/gata1/GATA1_standard_hvg_pert_filtered.h5ad, variant and theta tables, and the processed array under allele/gata1_arrays.npz. Source: GEO GSE215253 and the PerturbNet repository.
• JAK1 (scSNV-seq): the genotyped single-cell object allele/jak1/sce_genotyped.rds, cell/gene/experiment metadata, theta table and allele/jak1_arrays.npz. Sources: ENA PRJEB48915 and Zenodo record 10418435.
• allele/joint_arrays.npz, allele/real_deltas.npz and allele/barcode_rebuild_provenance.json: compact processed products and provenance for the allele analyses.
• allele/esm1v_embeddings.npz: ESM-1v variant embeddings used as molecular features. Model weights are not redistributed; cite and follow the terms of FAIR ESM and Meier et al. (2021).

Broader perturbation panel

The following paper-used resources are included when their complete source resource is below the 100 GB release threshold:

• external/scperturb/: selected scPerturb h5ad objects for Norman, Replogle K562 essential, Adamson, McFarland, Frangieh RNA/protein, Papalexi ECCITE RNA/protein and sciPlex3, together with the source manifest. Source: scPerturb Zenodo record 10044268.
• external/vcc/adata_Training.h5ad and its small validation/gene-name tables: the VCC training object used in the paper. Source: Arc Virtual Cell Atlas and the associated Cell paper (DOI 10.1016/j.cell.2025.06.008).
• external/gse306429/: the combined demultiplexed single-cell object and processed score/pseudobulk objects from GSE306429. Source: GEO GSE306429.
• external/perturbmulti/: processed CRISPR RNA and protein objects and the source README for PerturbMulti. Source: GEO GSE275483; the source README identifies the processed release as CC BY 4.0.

Files are stored in their native spaces. No results/ tables, differential- expression result objects, model checkpoints, prediction arrays, figures or analysis scripts are included.

Deliberate exclusions

Resources whose complete source resource exceeds 100 GB are not mirrored in this repository. Their provenance is retained here so that they can be retrieved from the original providers:

• Tahoe-100M full h5ad collection (about 320 GB),
• X-Atlas/Orion (about 118 GB),
• the KOLF pan-genome object (about 234 GB; see the Arc Virtual Cell Atlas release), and
• the Parse 10M PBMC object (about 212 GB; source/hosting details are given in the PertResolve manuscript and its preregistration).

The protocol-excluded non-cell-level CIGS and cpg0016 resources are also not included. These exclusions do not remove the corresponding benchmark definitions from the manuscript; they only avoid redistributing very large or out-of-scope source resources.

Provenance, licensing and citation

The Hugging Face repository is marked license: other because it combines author-generated annotations/derived products with files derived from public third-party datasets. This label is intentional and is not an Apache-2.0 grant for the third-party files. The original provider terms and attribution requirements apply to each source dataset; see NOTICE, the source manifests and per-resource source notes. The benchmark metadata and provenance text written for this release may be reused with attribution to the PertResolve authors, but no underlying third-party data are relicensed.

Please cite the PertResolve manuscript and the original source records listed above when using these data. MANIFEST.sha256 records the checksum of every included data or metadata file.