perturbmatch_analysis

Analysis scripts and processed data for reproducing figures in:

Yeung, J., Tan, J., et al. (2026). Joint analysis of multiply perturbed cells improves statistical power and cost efficiency in Perturb-seq. bioRxiv. https://doi.org/10.64898/2026.07.10.737863

License

• Code (scripts/, utils/, config.R): MIT License
• Data (data/): CC BY 4.0

Data availability

• GEO accession (pilot & sublib2): GSE337988 — raw and processed data
• Replogle et al. (2022) reanalysis: raw data at GSE146194; processed DE matrices at Zenodo (10.5281/zenodo.21232999)
• Processed figure data: included in data/ (CSV files for all figure panels)

Repository structure

├── config.R                    # Central path configuration
├── data/                       # Processed CSV data for all figure panels (~87 MB)
│   ├── fig1/                   # Figure 1: guides per cell (pilot)
│   ├── fig1_sup/               # Sup Figure 1: sprinter interaction effects
│   ├── fig2/                   # Figure 2: information loss + full vs marginal (pilot)
│   ├── fig2_sup/               # Sup Figure 2: DE method comparisons
│   ├── fig3/                   # Figure 3: statistical power (pilot)
│   ├── fig4/                   # Figure 4: sublib2 analysis
│   └── fig5/                   # Figure 5: Replogle reanalysis
├── scripts/                    # R scripts that generate each figure
│   ├── fig1/                   # 4 scripts → Fig 1B-G, Sup S1F
│   ├── fig1_sup/               # 1 script  → Sup S1D
│   ├── fig2/                   # 2 scripts → Fig 2B-G, Sup S3A
│   ├── fig2_sup/               # 2 scripts → Sup S2A-E
│   ├── fig3/                   # 1 script  → Fig 3B-C, Sup S3B
│   ├── fig4/                   # 4 scripts → Fig 4B-I, Sup S4A-E
│   └── fig5/                   # 3 scripts → Fig 5B-H, Sup S5A-D
└── utils/
    ├── crop_utils.R            # Plotting, info imbalance, SPRINTER screening
    ├── rms_cosine_theory.R     # Theoretical cosine similarity functions
    └── sc_limma_voom.R         # Single-cell limma-voom (RunScLimmaVoom, scVoom)

Quick start: reproduce figure panels

The data/ directory contains all processed CSV files needed to recreate every figure panel. Each CSV corresponds to a specific panel as documented in FIGURE_PANEL_DATA_MAP.md and SUPFIGURE_PANEL_DATA_MAP.md.

Running the full analysis pipeline

The scripts in scripts/ contain the complete analysis code that generates both the figures and the underlying CSV data from processed single-cell objects. To re-run the full pipeline:

1. Edit config.R to set DATA_ROOT_GSTORE and DATA_ROOT_SCRATCH to your local data paths
2. Ensure R dependencies are installed (see below)
3. Run individual scripts:

   Rscript scripts/fig1/00-show_nguides_across_MOI.R

R dependencies

CRAN/Bioconductor packages:

• dplyr, tidyr, ggplot2, data.table, Matrix, argparse
• ggrepel, ggrastr, lsa, viridis, pheatmap, ComplexHeatmap
• SingleCellExperiment, SummarizedExperiment
• limma, edgeR, scran, scrapper
• FNN, ggrepel, ggrastr, parallel, scales

For differential expression we use RunScLimmaVoom (provided in utils/sc_limma_voom.R), a limma-voom variant that replaces the standard lowess mean-variance trend with a robust regularized loess fit via scrapper::fitVarianceTrend and uses scran size factors instead of CPM normalization. Additional utility functions are provided in utils/crop_utils.R (scatter plots, information imbalance, SPRINTER interaction screening).

All dependencies are available from CRAN or Bioconductor.

Figure-to-script mapping

Figure	Panels	Script
Fig 1	B, C, D	scripts/fig1/00-show_nguides_across_MOI.R
Fig 1	E, F	scripts/fig1/2-downstream_limma_full_covar_with_heatmap.R
Fig 1	G	scripts/fig1/extract_go_msigdb_csv_from_rds.R
Fig 2	B, C	scripts/fig2/1-compare_full_regression_vs_marginal_with_barplots.R
Fig 2	D, E, G	scripts/fig2/10-calculate_information_loss_pilot_cleaned.R
Fig 3	B, C	scripts/fig3/20-calculate_statistical_power_pilot.R
Fig 4	B	scripts/fig4/00-nguides_by_MOI.R
Fig 4	C	scripts/fig4/10-calculate_information_loss_sublib2.R
Fig 4	D, E, F	scripts/fig4/20-calculate_statistical_power_sublib2.R
Fig 4	G, H, I	scripts/fig4/30-calculate_SNR_vs_cosine_similarity_sublib2.R
Fig 5	B, C	scripts/fig5/00-show_nguides_in_Replogle.R
Fig 5	D-H	scripts/fig5/20-calculate_statistical_power_Replogle_top_3000.R
Sup S1	A, B, C	scripts/fig1/3-correlate_residuals_sprinter_limma_spline.R
Sup S1	D	scripts/fig1_sup/3-correlate_residuals_sprinter_limma_spline_refit_top_covar_check_LFCs.R
Sup S1	F	scripts/fig1/00-show_nguides_across_MOI.R
Sup S2	A	scripts/fig2/1-compare_full_regression_vs_marginal_with_barplots.R
Sup S2	B, E	scripts/fig2_sup/4-compare_LimmaVoom_vs_other_methods_by_MOI.R
Sup S2	C, D	scripts/fig2_sup/4-FRPerturb_and_Othersdownstream_2026.R
Sup S3	A	scripts/fig2/10-calculate_information_loss_pilot_cleaned.R
Sup S3	B	scripts/fig3/20-calculate_statistical_power_pilot.R
Sup S4	A, B, C	scripts/fig4/00-nguides_by_MOI.R
Sup S4	D, E	scripts/fig4/30-calculate_SNR_vs_cosine_similarity_sublib2.R
Sup S5	A, B, C	scripts/fig5/20-calculate_statistical_power_Replogle_top_3000.R
Sup S5	D	scripts/fig5/00-ntargets_by_perturbation_replogle.R

Panels 1A, 2A, 2F, 3A, 4A, 5A are schematics created in Adobe Illustrator (no script).

Related resources

• perturbmatch package: https://github.com/Genentech/perturbmatch -- the standalone R package for propensity-score matching in high-MOI Perturb-seq experiments.