Perturbation-induced gene programs map the hidden regulatory architecture of complex diseases

Perturbscape learns perturbation-induced gene programs from 15,859 CRISPR Perturb-seq profiles, boosting heritability informativeness up to 4.3-fold versus steady-state prioritization across 53 complex diseases and traits.

• Why it matters: Steady-state molecular atlases miss regulatory programs that appear only after cellular perturbation, limiting how well genetic variation can be mapped to disease mechanisms.
• What they did: The framework uses linear and nonlinear contrastive learning plus program-learning methods to model multiscale programs per perturbation, then integrates them with human genetics using mean GWAS N=291K.
• The result: Perturbation-specific programs retain significant conditional heritability informativeness across 34 matched cell system–trait comparisons and yield 2.5–5.3-fold higher trait relevance scores than control-derived or pooled summaries.

AI summary of the abstract · may contain errors · not medical advice · read the paper.