Perturb-ME: Scalable mechanism discovery from phenotype-enriched genome-wide screens

In plain English

The preprint enriches melanoma cells with high or low MHC-I surface expression after a genome-wide CRISPR screen, then measures RNA, protein, and guide identity in the selected cells. Its regulatory analysis connects 221 regulators to 1,998 responsive genes, seven co-functional modules, and nine programs; an AI co-scientist then proposes biological interpretations for those modules.

What they found

• The analysis recovered the canonical interferon-gamma–MHC-I program and organized regulators into seven co-functional modules and nine programs.
• The AI co-scientist linked modules to trafficking, proteostasis, and chromatin mechanisms that remain hypotheses for further testing.

The catch

• The work is a preprint and has not completed peer review.
• The demonstration centers on one phenotype in a melanoma-cell context.
• Agent-generated module interpretations are author-reported hypotheses, not independent experimental confirmation.

Methods at a glance

Study type

Genome-scale CRISPR perturbation screen with phenotype enrichment and multimodal single-cell profiling

Biological system

MHC-I surface regulation in melanoma cells

Organism

Human

Dataset

Phenotype-enriched genome-wide CRISPR screen with RNA, protein, and guide readouts

Sample size

The abstract reports 221 regulators and 1,998 responsive genes; total profiled cell count is not reported

Model architecture

Regulatory modeling followed by an AI co-scientist for module interpretation

Training objective

Infer regulatory relationships, co-functional modules, and responsive gene programs

Code

Code is available at https://github.com/Genentech/PerturbME.

Data

The authors link a Figshare record at https://doi.org/10.6084/m9.figshare.33264225.

Tasks and metrics

Tasks: regulator-to-gene mapping; co-functional module discovery; gene-program discovery; mechanism hypothesis generation

Metrics: responsive-gene count; module coherence; program recovery

Baselines: Baseline details are not reported in the abstract

Principal findings

• The model linked 221 regulators to 1,998 responsive genes.
• The analysis reported seven co-functional modules and nine gene programs.
• The workflow recovered the canonical interferon-gamma–MHC-I pathway.

What remains unresolved

External validation: Not reported

• The preprint has not completed peer review.
• The reported application is limited to MHC-I regulation in one melanoma-cell setting.
• AI-generated mechanism interpretations require targeted experimental validation.