This project does not redistribute the scPerturb h5ad files, Geneformer weights, or large path-D caches in Git.
Download the harmonized scPerturb h5ad files and place them under data/:
| Dataset | Expected local path | Use in manuscript | Required contents |
|---|---|---|---|
| Frangieh et al. 2021 melanoma/TIL Perturb-CITE-seq | data/FrangiehIzar2021_RNA.h5ad |
Frangieh CCND1 worked example, token coverage, library-size contamination, DESeq2 target | raw counts, perturbation labels, sgRNA/replicate metadata, gene symbols and Ensembl IDs |
| Replogle et al. 2022 K562 essential-scale CRISPRi | data/ReplogleWeissman2022_K562_essential.h5ad |
cross-dataset DESeq2 increment tests, GATA1 state-shift/de-circularization | raw counts, perturbation labels, gemgroup/batch metadata, gene symbols and Ensembl IDs |
Access: see scPerturb (Peidli et al. 2024) and https://scperturb.org.
The repository expects Geneformer V2-104M to be available locally:
Geneformer/Geneformer-V2-104M/
Geneformer/geneformer/token_dictionary_gc104M.pkl
Geneformer source/weights are third-party assets and are not redistributed here. Install Geneformer separately so the geneformer Python package is importable.
Large generated artifacts are not intended for Git. They are available on Zenodo at https://doi.org/10.5281/zenodo.20729460, including:
| Artifact class | Example local path | Purpose |
|---|---|---|
| path-D caches | benchmark_output/pathd_cache/CCND1_cache.npz |
cached per-cell/per-gene responses enabling NumPy-only resampling and readout tests |
| DESeq2 outputs | benchmark_output/deseq2/*_replogle_deseq2.npz |
primary target and increment-test outputs |
| validation outputs | benchmark_output/pathd_validation.csv |
path-D vs official engine validation |
The Zenodo bundle restores the cached intermediate files under benchmark_output/ when extracted at the repository root.
The small curated figure-input CSVs — results/data/token_coverage.csv, results/data/CCND1_curve_main.csv, results/data/CCND1_curve_abl.csv — are committed to this repository (not Zenodo), so Figure 3 reproduces from a clean checkout.
The full path-D cache generation requires running Geneformer forward passes for WT and KO token sequences. A CUDA GPU is recommended; the worked Frangieh CCND1 path-D cache was generated on a 4 GB NVIDIA GPU by batching/subsampling.
Representative commands:
python scripts/pipeline/cache_pathd.py CCND1 data/FrangiehIzar2021_RNA.h5ad
python scripts/pipeline/validate_pathd.py
python scripts/pipeline/resample_pathd.py CCND1
python scripts/diagnostics/deseq2_groundtruth.py
python scripts/diagnostics/deseq2_replogle.py HSPA9
python scripts/diagnostics/deseq2_replogle.py SUPT6H
python scripts/diagnostics/deseq2_replogle.py CSE1L
python scripts/diagnostics/deseq2_replogle.py DHX15
python scripts/diagnostics/deseq2_replogle.py GATA1See README_reproduce_figures.md for panel-level figure provenance and reported_numbers_source_table.csv for manuscript-number provenance.