nicolas-lynn/vcc-perturb

Virtual Cell Challenge 2025 H1 hESC training-set perturbation atlas

CRISPRi (gene knockdown) in H1 human embryonic stem cells. Single-cell expression in log-normalized counts.

Generated 2026-06-08 as one of three companion atlases (Norman, Replogle, VCC).

File schema (each config / single-config repo)

File	Shape	Description
pseudobulks.h5ad	(50, n_genes)	50 control pseudobulks (15 cells each, log-normalized means). Cell-type-specific baseline.
coexpression.h5ad	(n_genes, n_genes)	Pearson coexpression across all cells in the source experiment (perturbed + control). Gene names on both axes.
de_log2fc.parquet	(n_perts, n_genes)	log2 fold change of perturbation vs control pseudobulk means. For Replogle (z-scored input), values are in z-difference units.
de_pvalue.parquet	(n_perts, n_genes)	Raw p-value per gene. DESeq2 Wald test for log-normalized inputs (Norman, VCC); Welch's t-test for z-scored input (Replogle).
de_padj.parquet	(n_perts, n_genes)	BH-adjusted p-value (FDR).
perturbation_deltas.parquet	(n_perts, n_genes)	Mean-difference effect (no cell-level QC filter). Useful as a baseline against the DE-filtered effect.
perturbation_zscores.parquet	(n_perts, n_genes)	Standardized effect (delta / control pseudobulk SD per gene, with adaptive floor). Comparable across atlases regardless of native units.
retained_perturbations_de.csv	(n_perts, ~15)	Per-perturbation metadata: n_cells_total, n_cells_passing target QC, target_log2fc, target_padj, n_DE counts at padj < 0.05 and 0.01.
retained_perturbations.csv	(n_perts, ~12)	Per-perturbation summary from the unfiltered delta pipeline (no per-cell target QC). Includes uniqueness, is_distinct_hit.
all_candidates_scored.csv	every scored candidate	For threshold tuning.
manifest.json, de_manifest.json	small	Build config + timestamp.

Generation pipeline

For each perturbation in the source experiment:

1. Per-cell target QC filter — keep only cells in which the target gene moved at least 50% in the expected direction:
   • Raw-count datasets (Norman, VCC): cell_norm_target >= 1.5 * mean_ctrl_norm_target (CRISPRa) or <= 0.5 * mean_ctrl_norm_target (CRISPRi).
   • z-scored data (Replogle): cell_target_z <= -1.0 (≥ 1 SD below pop mean).
   • Combinatorial perturbations require all listed targets to pass.
2. Drop perturbations with fewer than 15 cells passing the filter.
3. Build 50 perturbation pseudobulks of 15 cells each from the filtered pool (sum of counts for raw-count datasets; mean of z for z-scored).
4. Build 50 control pseudobulks the same way (no target filter).
5. Test DE per gene:
   • DESeq2 (negative-binomial Wald test, via pyDESeq2) on count pseudobulks.
   • Welch's t-test on z-scored pseudobulk means.
6. BH-adjust the per-perturbation p-values to give de_padj.

Coexpression matrix is Pearson correlation across all cells in the source h5ad (perturbed + control combined), after log-normalization (raw-count sources) or directly on z-values (Replogle).

Caveats

• Pseudobulks are random subsamples from one cell pool, not independent biological replicates. P-values are anticonservative for strict FDR claims. The effect sizes (log2FC / z-difference) and DE-gene rankings are robust; use those for downstream modeling and treat p-values as a ranking aid rather than a calibrated false-discovery rate.
• The per-cell target QC filter selects the subpopulation in which the perturbation strongly landed. That subpopulation typically also has a broader cell-state shift, so DE counts will exceed what you would see on the unfiltered population. Compare perturbation_deltas (no filter) against de_log2fc (filtered) to gauge that bias.
• Native effect units differ across atlases (log2FC for Norman/VCC; z-difference for Replogle). perturbation_zscores.parquet is the unit-free standardization (delta / control pseudobulk SD).

Source

Arc Institute Virtual Cell Challenge 2025 training set. gs://arc-institute-virtual-cell-atlas/virtual-cell-challenge/2025/train/adata_Training.h5ad

Citation

If you use this dataset, please also cite the source paper above.