Reproducible computational vignettes — CRISPR-perturbation multi-omics identifiability

Code and result files that regenerate the illustrative computational analyses in the review:

Asediya, V. When can CRISPR-perturbation multi-omics support a causal claim? A method-selection framework for livestock functional genomics. (Under review, Briefings in Bioinformatics.)

All analyses run on synthetic or public data; no new experimental data were generated.

---

Contents

• What this is
• Study → figure map
• Repository layout
• Installation
• Reproduce the figures
• Data
• Scope and integrity
• Citation
• Licence
• Contact

What this is

This repository accompanies a methods review, not a primary benchmarking study. The computations here are small, clearly-labelled illustrative reproductions that calibrate and visualise the review's framework; each reproduces or makes concrete a result already established in the cited literature. The substantive evidence in the paper is the published benchmark record, which the review synthesises.

Study → figure map

Study	Script	Backs in the paper	Status
A	study_A_identifiability.py	Main Fig 2; Supp Fig S1a–b	load-bearing
B	study_B_norman_benchmark.py	Main Fig 9a	load-bearing
C	study_C_grand_benchmark.py	Supp Fig S1c (contract matrix)	load-bearing
D	study_D_obstruction_hierarchy.py	Supp Fig S1d	load-bearing
F	study_F_synergy.py	Main Fig 9b	load-bearing
G	study_G_esm.py	ESM-2 constraint ↔ conservation (support for Figs 11–12)	load-bearing
H, H2, I, J	study_H*.py, study_I*.py, study_J*.py	Supp Fig S1e–f + editability material	exploratory
K, L	unrun_scaffolds/	—	not run (see unrun_scaffolds/NOT_RUN.md)

study_figures.py regenerates study_validation_causal.png and study_perturbation_ml.png from the computed result files.

Repository layout

Path	Purpose
study_*.py	One analysis ("study") each; see the map above
study_*_results*.json	Machine-written numeric results for each completed study (figures are built from these)
study_figures.py	Builds the publication figures from the result JSONs
prepare_norman_cache.py	One-time cache of the Norman 2019 pseudobulk effects → norman_cache.npz
study_validation_causal.png, study_perturbation_ml.png	Rendered figure outputs
unrun_scaffolds/	Runnable scaffolds not executed for this paper (no conclusions depend on them)
requirements.txt	Python dependencies
.zenodo.json, CITATION.cff	Archive + citation metadata
LICENSE	MIT licence

Installation

Python 3.9+ is recommended.

python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt

Reproduce the figures

# 1. (once) fetch the public Norman 2019 dataset (see Data) and build its cache
python prepare_norman_cache.py

# 2. regenerate the figure outputs from the computed results
python study_figures.py

The synthetic-network studies (A, C, D) generate their own data in-script from fixed random seeds, so they are fully self-contained.

Data

The Norman et al. (2019) combinatorial Perturb-seq dataset is public and large, so it is not redistributed here. Fetch it from its primary source (GEO GSE133344 / the authors' release) and build the analysis cache with prepare_norman_cache.py. The ESMFold / ESM-2 analyses (Studies G, J) use the public ESM Atlas / fair-esm weights. No new biological data underlie any result here.

Scope and integrity

• Illustrative, not primary research. The vignettes calibrate the framework; they are not a comprehensive method benchmark.
• The Norman 2019 reproduction is run on the Norman 2019 dataset only — it is not a multi-dataset or "whole-corpus" benchmark.
• Exploratory studies (H, H2, I, J) back the exploratory supplementary panels and are offered as illustration, not validated methods, exactly as stated in the paper.
• The identifiability atlas is regenerable, not shipped. The exploratory oracle panel (Supplementary Fig S1e, "≈87,000 algorithm runs") is built from Study H's atlas study_H_atlas.json, which is large and intentionally excluded (see .gitignore). Regenerate it with python study_H_identifiability_atlas.py, then python study_H2_oracle.py; the deposited study_H2_oracle.json is its downstream summary.
• Main Figure 2 is an illustrative schematic of a single network instance (hand-composed); the quantitative, multi-seed version of that analysis is Supplementary Fig S1, produced by Study A.
• Not run for this paper (unrun_scaffolds/). study_K_* was executed on the Norman 2019 dataset only; study_L_*'s foundation-model arm was never run (it needs a GPU and an unimplemented foundation_predict()). Their machine-written result JSONs are intentionally omitted so they cannot misrepresent scope. No conclusion in the manuscript depends on Study K or L.

Citation

If you use this code, please cite the article and the software archive (a CITATION.cff is included, so GitHub's "Cite this repository" button works):

Asediya, V. When can CRISPR-perturbation multi-omics support a causal claim? A method-selection framework for livestock functional genomics. (Under review, Briefings in Bioinformatics.)

The software archive carries a Zenodo DOI (badge above); the article DOI will be added on publication.

Licence

Code is released under the MIT Licence (see LICENSE). The manuscript text and figures are © the author.

Contact

Varunkumar Asediya — Institute of Veterinary Medicine, Nicolaus Copernicus University in Toruń, Poland — varun@doktorant.umk.pl — ORCID 0000-0002-9402-0349