Perturb-ME

Analysis code for "Perturb-ME: Scalable mechanism discovery from phenotype-enriched genome-wide screens" (Wang, Gu, Frangieh et al.), bioRxiv (2026), doi.org/10.64898/2026.08.17.745330.

Perturb-ME combines genome-scale CRISPR screening, marker-based phenotypic enrichment by FACS, and multimodal single-cell profiling by Perturb-CITE-seq. Applied to surface MHC-I (HLA-A/B/C) in the A375 melanoma line, it profiles HLA-low and HLA-high cells with matched RNA, surface-protein and sgRNA measurements, fits a regularised linear model of perturbation effects, and decomposes that model into co-functional regulatory modules and gene programs.

Data

This repository contains code only. The data live in two places:

• Raw and processed sequencing data — NCBI Gene Expression Omnibus, accession in the paper's Data Availability statement. Single-cell gene expression, CITE-seq, cell hashing and CROP-seq guide-capture libraries from 20 10x channels, plus bulk sgRNA enrichment sequencing from genomic DNA.
• Analysis-ready objects — figshare, doi.org/10.6084/m9.figshare.33264225: the two AnnData objects (353,091 cells and the 277,152 sgRNA-singlet subset), the EM-refined ElasticNet coefficient matrix, the target-module and gene-program partitions, MAGeCK outputs, and the Co-Scientist prompt and response.

Layout

src/            66 analysis modules, organised by stage
  RNA/          cell and gene filtering, QC
  Hash/         hashtag demultiplexing (Cumulus)
  CITE/         antibody-derived tag normalisation
  CROP/         guide calling and per-target coverage
    linear_model/   lm_target_EM.py         the published ElasticNet + EM model
                    cluster_by_kmeans*.py   target modules and gene programs
  gDNA/         bulk sgRNA enrichment, MAGeCK inputs
  clustering/   condition and expression clustering, cell-cycle scoring
  manuscript_figures/   one module per published panel, plus source-data builders
scripts/        batch drivers and the data-deposit builders
notebook/       5 notebooks that produced published panels

Figure panels map to src/manuscript_figures/ by name — 1C_cluster_conditions.py, 2E_cluster_by_kmeans_genome_wide_features.py, and so on. The Fig. 2b,c module and gene-program partition is produced by notebook/20260519_Figure2B_Rerun_BetaCorr.ipynb.

Environment

bash setup.sh          # creates the `perturbme` conda environment (Python 3.10)
conda activate perturbme

Core dependencies: scanpy, anndata, scikit-learn, scipy, pandas, harmonypy, scvi-tools. MAGeCK v0.5.9.2 is used for the bulk screen analysis and is invoked separately.

⚠️ Paths are institution-specific

This code was written to run on a single institutional HPC cluster and contains hardcoded absolute paths (/gstore/..., /gnet/..., and in the older linear-model modules /ahg/regevdata/...). They appear as module-level constants, usually named REPO, DATA or similar near the top of each file.

To run this elsewhere, point those constants at your own copy of the data from GEO/figshare. Nothing else about the code is site-specific. We have deliberately left the paths as they were when the analysis was run rather than rewriting them after the fact.

Citation

Wang, H., Gu, J., Frangieh, C. J., Cuoco, M. S., Zhao, M., Sett, A., Beyer, T., Pang, K., Stolte, E., Xu, Z., Leskovec, J., Rozenblatt-Rosen, O., Natarajan, V., Geiger-Schuller, K., Thakore, P. I. & Regev, A. Perturb-ME: Scalable mechanism discovery from phenotype-enriched genome-wide screens. bioRxiv (2026). doi:10.64898/2026.08.17.745330

License

MIT — see LICENSE.