Zafrens/zscreen_pilot_data

Z-Screen Source Data (June 2026 Release)

This deposit contains the curated public source-data layers for the Z-Screen platform, a modality-agnostic single-nanowell drug-discovery screen that couples transcriptomic counts, imaging morphology embeddings, and compound chemistry across a shared public compound identifier.

Each row of the primary data represents a single nanowell (90 microns diameter by 120 microns tall, holding approximately 2 to 10 cells), not a single cell.

• Nanowells profiled: 768,162 across 49 devices (numbered 0 to 48)
• Genes measured: 46,944 (transcriptomic counts)
• Cell lines / clone variants: 6 (HEK293, HEK293-clone, A549, H1650, THP1, AEC7)
• Libraries: 14 (named-control compounds, OBOC combinatorial libraries, control beads)
• Unique compounds / precursors: 142,227 (78 named controls, 142,149 OBOC building blocks, plus controls)

The primary tables are joined 1-to-1 across counts.h5ad, MasterFile.parquet, and the chemistry layer on the nanowell axis, and share a public_compound_id key that links biology to chemistry.

Contents

raw_counts_data/               Raw transcriptomic counts + master well table
normalized_image_embeddings/   Normalized morphology features + biomarker intensities
chemical_features/             Public compound / building-block chemistry features
active_seq_same_well/          Same-well RNA + imaging calibration layer (ActiveSeq)
external_reference/            External benchmark atlases (ranked gene lists) + gene sets
README.md                      This file
MANIFEST.csv                   Per-file SHA256, size, and parquet row/column counts

raw_counts_data/ (~3.9 GB)

File	Shape	Description
counts.h5ad	768,162 wells x 46,944 genes	Raw transcriptomic counts (AnnData / HDF5).
MasterFile.parquet	768,162 x 274	Master per-well table (well/device/library/compound/QC metadata), 1-to-1 with counts.h5ad.
data_shape_summary.md	-	Human-readable data map (devices, cell lines, libraries, nomenclature).
data_shape_visualization.html	-	Interactive overview of dataset structure.

On the Zenodo record, each top-level folder above is provided as a single .zip (raw_counts_data.zip is 4.0 GB, MD5 796235d30c4e7e821fe8453f07afae45). On Hugging Face the same files are provided uncompressed for direct/streamed access.

normalized_image_embeddings/ (~6 MB)

File	Shape	Description
normalized_image_embeddings.parquet	21,795 x 39	Compound-level normalized morphology (image) features, keyed by public_compound_id.
biomarker_intensity_layer.parquet	18,942 x 11	Protein/cellular-state marker intensities (BRD4, p21/CDKN1A, DAPI, actin, ConA, p62/SQSTM1).

chemical_features/ (~295 MB)

File	Shape	Description
compound_chemistry_feature_master.parquet	144,645 x 961	Compound + building-block chemistry, keyed by public_compound_id. Formula/InChIKey where available, plus Chemeleon and ChemBERTa embeddings (whole-molecule and per-building-block BB0 to BB4).

active_seq_same_well/ (~440 MB)

The public layer supporting the claim that RNA and morphology were measured from the same physical nanowell. The primary deliverable is self-contained; the remaining files are the full pre-join latent spaces so the same-well matching can be reproduced from scratch.

File	Shape	Role
joined_same_well_data_with_public_compound_id.parquet	18,727 x 550	Primary deliverable. Matched same-well rows carrying 32 RNA latent dims (D00-D31), 448 image latent dims (img_lat_*), well geometry, RNA QC, and public compound mapping.
activeseq_control_name_public_compound_map.parquet	36 x 7	Control-name to public_compound_id mapping (35 named compounds, 2 ActiveSeq devices).
imaging_lats.parquet	126,662 x 451	Full imaging-patch latent space (pre-join).
rnaseq_lats.parquet	61,101 x 33	Full RNA-cell latent space (pre-join).
assay_well_dataframe.parquet	126,662 x 35	Imaging-patch geometry / detection metadata.
rnaseq_obs.parquet	61,101 x 27	RNA-cell observation metadata / QC.
matches_table.parquet	20,222 x 8	Imaging-to-RNA well matches used to build the joined table.

Note: the redundant joined_same_well_data.parquet (a strict subset of the _with_public_compound_id table, differing only by a pandas index column) is intentionally excluded from this deposit.

external_reference/ (~182 MB)

Clean, reusable external-benchmark surface. The core assets are the ranked gene-list signature tables (*_rank_signatures_top_genes.parquet) plus their metadata, catalogs, provenance, and gene sets.

Atlas	Key ranked-gene-list table
Replogle 2022 K562 genome-wide CRISPR KO	crispr_perturbation_atlases/replogle_2022_k562_gwps/replogle_rank_signatures_top_genes.parquet
Replogle 2022 RPE1 CRISPR KO	crispr_perturbation_atlases/replogle_2022_rpe1/rpe1_rank_signatures_top_genes.parquet
Norman & Weissman 2019 K562 CRISPRa	crispr_perturbation_atlases/norman_weissman_2019_k562_crispra/norman_rank_signatures_top_genes.parquet
Selected THP1 scPerturb CRISPR	crispr_perturbation_atlases/scperturb_thp1_selected/scperturb_rank_signatures_top_genes.parquet
LINCS L1000 compound signatures	compound_atlases/lincs_l1000/processed/lincs_l1000_rank_signatures_top_genes.parquet
Tahoe-100M	compound_atlases/tahoe_100m/ (metadata + capmatinib slice)
Control transcriptome expansion (2026-05-25)	compound_atlases/control_transcriptome_expansion_20260525/

Also included: benchmark_outputs/ (prior LINCS vs Z-Screen overlap/correlation outputs), catalog/ (dataset catalog, file manifest with hashes, provenance), gene_sets/ (Enrichr KEGG / Hallmark / Reactome GMTs, disease-reversal and aging signature JSONs), and the provenance/fetch scripts documenting how the tables were produced.

Curation note: the raw external .h5ad matrices (scperturb_downloads/, ~6.3 GB) are not bundled here. They are publicly re-downloadable from scPerturb's own Zenodo record; see crispr_perturbation_atlases/scperturb_thp1_selected/source_manifest/ and external_reference/fetch_zenodo.py for retrieval. Only the processed rank-signature tables are included, so downstream comparisons do not require rebuilding them.

Loading

Parquet (Python):

import pandas as pd
master = pd.read_parquet("raw_counts_data/MasterFile.parquet")
chem   = pd.read_parquet("chemical_features/compound_chemistry_feature_master.parquet")
# link biology to chemistry on the shared key
merged = master.merge(chem, on="public_compound_id", how="left")

Counts (AnnData):

import anndata as ad
adata = ad.read_h5ad("raw_counts_data/counts.h5ad")   # 768,162 obs (nanowells) x 46,944 vars (genes)

Hugging Face datasets (streaming, no full download) - the main parquet tables are exposed as named configs for the dataset viewer and load_dataset:

from datasets import load_dataset
ds = load_dataset("Zafrens/zscreen_pilot_data", "master_well_table", split="train", streaming=True)
print(next(iter(ds)))

Available configs: master_well_table, chemical_features, image_embeddings, biomarker_intensity, same_well_rna_image, activeseq_compound_map, ref_replogle_k562, ref_replogle_rpe1, ref_norman_k562, ref_scperturb_thp1, ref_lincs_l1000. The counts.h5ad matrix is not exposed to the viewer (load it with AnnData as shown above); all other files remain directly downloadable.

Data fields

Key columns of the headline tables (public_compound_id is the shared join key across all layers).

master_well_table (raw_counts_data/MasterFile.parquet, 768,162 x 274) - one row per nanowell, aligned 1-to-1 with counts.h5ad:

• Identifiers: obs_id, device_id, sample_id, cell_line, zlibrary, condition_role
• Compound / chemistry: public_compound_id (+ _source), public_bb0_id..public_bb4_id, chemistry_grain, name, molecular_formula, inchi_key
• Chemistry embedding: whole_chemeleon_chemeleon_proj_000.._NNN (whole-molecule Chemeleon projection)

chemical_features (144,645 x 961) - one row per public_compound_id:

• Support / provenance: source_*, condition_*, token_single_*, token_all_*, public_primary_bb0_id..bb4_id, molecular_formula, inchi_key
• Embeddings: whole-molecule Chemeleon and ChemBERTa feature blocks, plus per-building-block Chemeleon blocks for BB0 through BB4

image_embeddings (21,795 x 39) - compound-level morphology:

• Keys: aggregate_id, public_compound_id, zlibrary, image_dataset, is_control, cell_line_context, public_condition_level_mrna_join_key
• Features: pca_00..pca_31 (32 principal-component image features)

biomarker_intensity (18,942 x 11) - marker-state intensities (*_channel_masked_mean_zrobust_mean): BRD4_anti_target_curated, CDKN1A_p21, DAPI, PHALLOIDIN_ACTIN, CONCANAVALIN_A, SQSTM1_p62.

same_well_rna_image (joined_same_well_data_with_public_compound_id.parquet, 18,727 x 550) - matched same-nanowell rows: D00..D31 (32 RNA latent dims), img_lat_* (448 image latent dims), well geometry, RNA QC metrics, and the public compound mapping columns.

ref_* tables - external ranked gene-list signatures; each row is a (signature, gene) rank entry used for gene-rank / phenomimic comparison. Treat a high similarity as a perturbation-state hypothesis, not direct target identification.

Integrity

MANIFEST.csv lists every file with its SHA256 checksum, byte size, and (for parquet) row and column counts. Verify a download with, for example, sha256sum -c after reformatting, or by recomputing hashes and comparing against the manifest.

Excluded from this release

The following were deliberately left out of the deposit: the Manuscripts_June2026 folder, the zscreen_v2 / v3 / v4 working directories, the raw external scperturb_downloads/ H5AD matrices, and the redundant active_seq_same_well/joined_same_well_data.parquet.

License and citation

License: CC BY-NC 4.0 (Creative Commons Attribution-NonCommercial 4.0 International). You may share and adapt the data with attribution, for non-commercial purposes only.

This dataset is archived on Zenodo. Cite the concept DOI, which always resolves to the latest version: 10.5281/zenodo.21195738. To pin the exact files of a specific release for reproducibility, use that release's version DOI instead (v1, 2026.06 = 10.5281/zenodo.21195739). Please also cite the associated Z-Screen manuscript once available.

@dataset{vijayan_zscreen_2026,
  author    = {Vijayan, Swamy},
  title     = {{Z-Screen Source Data: single-nanowell transcriptomics,
               image embeddings, and compound chemistry}},
  year      = {2026},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.21195738},
  url       = {https://doi.org/10.5281/zenodo.21195738}
}