Datasets:

KiritSingh
/

spot-CD4-Marson

license: mit
pretty_name: spot Stage 1 — CD4 continuous transcriptional-program scores
tags:
  - single-cell
  - immunology
  - perturb-seq
  - CD4-T-cells
  - transcriptomics
  - reproducibility
size_categories:
  - 100K<n<1M
source_datasets:
  - CZI Virtual Cells Platform — Primary Human CD4+ T Cell Perturb-seq
language:
  - en
annotations_creators:
  - machine-generated

spot Stage 1 — CD4 continuous transcriptional-program scores

This repository contains the public data release for spot Stage 1, method version stage1-continuous-v3.0.1. It provides continuous RNA-program scores for a quota-balanced non-targeting-control (NTC) subset of the Marson GWCD4i primary-human-CD4 Perturb-seq dataset.

The scores are measurements of specified transcriptional modules. They are not categorical cell-type, lineage, protein, cytotoxicity, or suppressive-function calls. The Stage-1 scoring and validation outputs contain no inferential p-values, q-values, or FDR estimates and do not report cell-type prevalence.

This dataset repository deliberately contains Stage 1 only. Stage 2–4 result artifacts use additional sources and are not bundled here.

Source and subset

Upstream public source:

Primary Human CD4+ T Cell Perturb-seq, CZI Virtual Cells Platform, v1.0, released 22 December 2025: https://virtualcellmodels.cziscience.com/dataset/genome-scale-tcell-perturb-seq
Zhu, Dann, et al., 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, https://doi.org/10.64898/2025.12.23.696273.
Authors' analysis repository: https://github.com/emdann/GWT_perturbseq_analysis_2025.

ntc_clustered.h5ad is a spot-derived pinned input object, not an original authors' release file. It contains 396,000 NTC cells: 33,000 cells from each of four coded donors in each of Rest, Stim8hr, and Stim48hr. This equal quota is a spot sampling decision and does not preserve the source population's condition proportions.

The object is 396,000 × 18,130. Its .X matrix is already normalized to a per-cell target total of approximately 9,819 and log1p transformed. No raw counts layer is retained. Cluster metadata are retained, but latent and UMAP matrices are not stored in the H5AD (obsm, obsp, layers, and uns are empty). Frozen display coordinates are supplied separately.

Scoring method

For cell c and program p:

score(c, p) = mean(.X[c, measured panel genes])
              - mean(.X[c, frozen expression-matched control genes])

The deterministic control construction is fully materialized in the release:

25 expression bins computed over the pinned 396,000-cell matrix;
eligible pool = detected genes minus every program marker and activation predictor marker;
program-order-invariant keyed SHA-256 selection;
50 controls per occupied marker bin, without replacement;
master seed 12345.

There are 11 primary axes: Th1-like, Th2-like, Th17-like, Tfh-like, Treg-like, CD4 CTL-like, Th9-like, Naive, Activated, Memory, and Checkpoint. The activation-adjusted CD4 CTL-like residual is a sensitivity/display lane and is not a primary pole.

Marker panels are curated canonical-marker panels with per-gene primary-source provenance in the registry. Masopust et al., https://doi.org/10.1038/s41577-025-01238-2, is used as a nomenclature framework, not as the source of the marker panels.

Files

Path	Purpose
`ntc_clustered.h5ad`	Pinned 396,000-cell normalized NTC scoring source object.
`stage1-continuous-v3.0.1/data/stage01_scores_full.parquet`	Authoritative 396,000 × 15 table: barcode, donor, condition, 11 primary scores, and one activation-adjusted sensitivity score.
`stage1-continuous-v3.0.1/data/stage01_summary_v3.json`	Full-table program summaries by condition.
`stage1-continuous-v3.0.1/data/stage01_umap_coordinates.json`	Frozen coordinates for the 40,000-cell display sample.
`stage1-continuous-v3.0.1/data/stage01_umap_overlay_v3_hf.json`	The same 40,000 barcodes with v3 scores; display-only.
`stage1-continuous-v3.0.1/method/`	Input identity, registry and marker provenance, frozen bins, controls, eligible pool, activation-association diagnostic, and Stage-2 scorer projection.
`stage1-continuous-v3.0.1/validation/`	Frozen gate specification, descriptive validation, semantic amendment, constituent evidence, and non-gating diagnostics.
`stage1-continuous-v3.0.1/contracts/`	Generic Stage-1 selection-contract schema.
`stage1-continuous-v3.0.1/receipts/`	Independent reconstruction receipt.
`MANIFEST.json`	Exact size and SHA-256 ledger for every current release file.

stage01_stage2_registry_view.json is only a Stage-1 scorer projection for downstream binding; it is not a Stage-2 result.

Validation status

Full-score raw SHA-256: de63b496e8121c77babe380e0c3b5ddfd66f9ce67d0d4e80f55645d177e27e5f.
Canonical full-score content SHA-256: 43c4296d5166740c334441a69df23bb440a073382bbe79628a3bb89e43d51316.
The independent reconstruction receipt reports all checks passing, including exact overlay-to-full score equality for all 40,000 display barcodes.
The 40,000-cell overlay is not the scientific analysis universe. An archived representativeness diagnostic contains one failed overlay-distribution check; use the full 396,000-cell table for analysis and summaries.
The frozen validation archive records that 0/33 program-condition pairs cleared its prospective small-panel leave-one-marker robustness gate. That result is retained for auditability, has active_gate:false in the current selection contract, and is not a claim that the programs are biologically invalid. Null/PENDING selectability fields inside the scorer registry are the pre-validation snapshot; stage01_selectability_v3.json records the later historical disposition.
stage01_activation_association_v1.json reports descriptive activation association for every primary axis and the pooled activation-adjusted CTL residual. Activation/timepoint association is not claimed to be removed.

External-dataset confirmation and protein, lineage, and functional validation remain separate biological work; they are not implied by this scoring release.

The HF overlay differs from the repository-generated overlay only in its top-level explanatory note. Cell records, coordinates, score fields, canonical score hash, and coordinate hash are unchanged. Both source and emitted raw hashes are recorded in MANIFEST.json; raw-byte reproduction is not claimed for that note-only packaging transform.

Reproducibility

The pinned input identity is:

source HF revision: e5fcf98b56a9302921d402e97fc5a190bd88f9a6;
ntc_clustered.h5ad raw SHA-256: 2edc6d318415c8b0ee779d707ab86e26ddb6f0274db51ab4a12f21ebfda50e43.

Reproduction and verification code is in the public spot repository. The input manifest, control algorithm, materialized controls/bins, validation constituents, and independent receipt are included here so the release can be audited without relying on moving branch names.

Privacy

The H5AD contains coded donor, guide, lane, and library identifiers but no donor demographics. In particular, it does not include age, sex, ethnicity, weight, height, smoking status, blood type, or collection date. Coded identifiers may be joinable to upstream public metadata and therefore should not be described as anonymous. Release files were scanned for machine-local paths and credentials.

License and release history

The official CZI dataset page declares the upstream dataset MIT License. The exact upstream copyright-holder notice is not supplied on that page, so this repository does not invent one. Spot-authored transformations and metadata are MIT licensed; see LICENSE and NOTICE.

The historical v2 release is preserved at immutable commit e5fcf98b56a9302921d402e97fc5a190bd88f9a6 and release tag stage1-continuous-v2. Its root stage01_umap_seed.json is intentionally absent from the current tree. v3.0.1 supersedes the earlier categorical-call, permutation-FDR, prevalence, paper-exact-embedding, and CP10k-equivalence claims; none of those claims is made by this release.

Current release tag: stage1-continuous-v3.0.1.