bendidiihab/vcc-2026-data

VCC 2026 — transformed sources and the H1 evaluation panel

The prepared data behind github.com/IhabBendidi/vcc-2026, an entry to the Virtual Cell Challenge 2026.

The task: predict how a cell responds to a CRISPRi knockdown in a cell line the model has never seen. Given only non-targeting control profiles for three anonymised contexts and a list of 300 target genes, emit 360,000 cells × 18,533 genes of raw counts.

This repo holds 46 files, 28.54 GB: the per-source screens already transformed onto the challenge's axes, and the H1 hESC panel used for local evaluation. It is the starting point of the pipeline, not raw data — the ~153 GB of upstream screens are not here.

⚠️ This is not a ready-to-train dataset. It is a set of intermediate artefacts addressed by hard-coded relative paths in the companion repo's scripts. The dataset viewer is disabled because nothing here is a table of examples. Use scripts/fetch_data.py, which reconstructs the expected directory layout and sha256-checks every file.

Getting it

git clone git@github.com:IhabBendidi/vcc-2026.git && cd vcc-2026
uv venv --python 3.12 .venv
uv pip install --python .venv/bin/python -r requirements.txt

hf auth login
.venv/bin/python scripts/fetch_data.py --target minimal

--target	size	what it buys
minimal	3.6 GB	re-run the shipped recipes end to end
hp046	27.3 GB	+ re-pool the challenge sources with different --sources / --rank-j
h1	27.4 GB	+ the same for the H1 panel
all	28.5 GB	everything

The difference is one file: sources/celldist/full_atlas.pkl (24.8 GB), read only when re-deriving a pool from scratch.

Structure

sources/ — 18 files, 26.5 GB

Per-source CRISPRi screens aligned onto the challenge's gene and perturbation axes and reduced to log2 fold changes.

file	size	what it is
celldist/full_atlas.pkl	24.8 GB	genome-wide per-source log2FC matrices, {source: (matrix, perts, genes)}. Needed only for the rank-j per-source subspace removal
celldist/nearest_logfc.pkl	485 MB	44 source keys on the H1 255-perturbation panel
celldist/vcc_logfc.pkl	420 MB	21 source keys on the challenge 300-perturbation panel. Schema {source: (dict{pert → log2FC vector}, gene_list)}
context_{A,B,C}.h5ad	662 MB	the challenge's own control cells, 18,400 × 18,533 raw counts each
conf_vcc_{A,B,C}.npz	34 MB	per-(perturbation, gene) confidence maps driving emission dispersion
prof_hp039_a080.h5ad	64 MB	the pooled profile, 900 × 18,533 CPM — the input to the winning transform
prof_hp046_snr.h5ad	64 MB	the same after the per-gene SNR soft-threshold — the profile actually submitted
pertrel/*.csv, depmap/*.csv, pert_energy.csv, *.npy	<1 MB	per-source reliability, DepMap essentiality, magnitude priors

The six sources pooled for the best submission are Feng_cpm, CD4_Tcell, K562_gwps, H1_2025, Heracles_HCT116, Heracles_HEK293T. The other keys are variants kept for ablations (_raw, _c2…_c20 are clip thresholds).

h1_panel/ — 28 files, 2.05 GB

The H1 human embryonic stem cell panel, built from Arc's public 2025 challenge release and binomially thinned to the 2026 sequencing depth (~20,000 UMI/cell). This is the only context where ground truth is held, so every local A/B decision was made here.

file	size	what it is
panel_real.h5ad	1.42 GB	ground truth: 255 perturbations × 400 cells + 18,400 controls
panel_controls_model.h5ad	225 MB	the disjoint model-visible control pool
panel_controls.h5ad	224 MB	the scoring-side controls
eval/cache_real/	108 MB	the scorer's cached DE tables — saves hours per run
eval/{baseline,anchor}/	140 KB	the 0-end and 1-end that define the scale
conf_*.npz	72 MB	seven confidence maps: the production one plus its ablation ladder and a shuffled null
arms_gab/prof/prof_gp_j15.h5ad	18 MB	the frozen best-arm profile

⚠️ panel_controls_model.h5ad is disjoint from panel_controls.h5ad on purpose. An earlier version of this rig let models emit cells drawn from the scoring reference; that leak was worth +0.02 of apparent score and every arm had to be re-run. Never fit on panel_real.h5ad.

Two conventions that are easy to get wrong

1. h1_ctrl_cpm.npy and ctrl_pseudobulk_cpm.npy are not the same kind of object. The first is a group-sum pseudobulk (colsum/total × 1e6, so a deep cell counts more); the second is the mean of per-cell CPM (every cell counts equally). On context A they differ by up to 518 CPM. scripts/make_ctrl_cpm.py rebuilds both and explains why.
2. scratch/ is the data root, not scratch. fetch_data.py reconstructs that tree because 372 scripts address data by relative paths under it.

What is deliberately not here

excluded	size	where to get it
raw upstream screens (Feng ×3, Replogle, Nadig ×2, K562 gwps, CD4, Heracles)	~153 GB	builders in the repo's scripts/
VCC 2025 H1 release	18 GB	public: gs://arc-institute-virtual-cell-atlas/virtual-cell-challenge/2025/
challenge controls bundle	632 MB	vcc datasets download controls
built submissions (.vcc)	8 GB	regenerable: sbatch scripts/hp046.sbatch

Licensing and provenance

Derived data. Every file here is a transformation of publicly released datasets. The transforms (axis alignment, log2FC reduction, pooling, thinning) are ours; the underlying measurements are not.

upstream	used as
Arc Institute Virtual Cell Challenge 2025 (H1 hESC)	the h1_panel/ and the H1_2025 pooled source
Arc Institute Virtual Cell Challenge 2026	context_{A,B,C}.h5ad
Replogle et al. 2022 (K562, RPE1 genome-scale CRISPRi)	K562_gwps, Replogle_RPE1
Nadig et al. 2025 (HepG2, Jurkat)	Nadig_*
Feng et al. (genome-wide fitness / non-fitness / targeted)	Feng*
Jiang et al. 2025 (six cancer lines)	Jiang_*
DepMap	the essentiality magnitude prior
Heracles (HCT116, HEK293T)	Heracles_*
CD4 T-cell CRISPRi screen	CD4_Tcell

Every source is publicly available. Nothing here derives from proprietary or unreleased data — the local copies we worked from are clones of public releases.

⚠️ license: other is deliberate. The upstream datasets are public but carry different terms from one another, and we do not relicense them. You are responsible for checking and honouring the original licence of each source you use, and for citing the original authors rather than this repo.

Why this dataset is private: the VCC 2026 competition is live (final round opens 2026-10-22), and these artefacts are the direct input to our submissions. It is a competitive hold, not a licensing one, and the intent is to open it after the competition closes.

No patient-level or personally identifying data is present; all content is cell-line or primary-cell transcriptomics reduced to per-gene summary statistics.

Reproducing the result

The companion repo's best submission scored overall 0.19820, rank 37 on the A/B/C validation panel. With --target hp046:

sbatch scripts/hp039.sbatch    # pool 6 sources  -> prof_hp039_a080.h5ad
sbatch scripts/hp046.sbatch    # SNR gate + emit -> HP046.vcc

The profile this produces is bit-identical to sources/prof_hp046_snr.h5ad, the one actually submitted.

Full findings, post-mortems and the job registry are in the GitHub repo — including why an earlier version of the same transform scored 0.15494 and what the 0.043 difference was.