Perturbation-induced gene programs fill a gap left by steady-state atlases, retaining independent genetic signal in all 34 matched cell–trait comparisons

Disease-gene ranking no longer rests only on steady-state expression, pathway or protein-network annotations; it comes directly from transcriptional programs that emerge after experimental perturbation. Earlier integrative approaches learned gene-to-disease weights mainly from precompiled steady-state multi-tissue features, which poorly cover regulatory programs that appear only under perturbation or in specific cellular, developmental or activation contexts. Across 15,859 perturbation-by-context profiles from 7 Perturb-seq systems and 53 diseases and traits (mean GWAS N = 290,733), perturbation-induced programs showed up to 4.3-fold higher heritability informativeness than steady-state multi-tissue gene prioritization and retained significant conditional heritability informativeness in all 34 matched cell system–trait comparisons after accounting for steady-state prioritization.

Each perturbation is first decomposed into a set of multiscale gene programs, then combined into a disease-weighted meta-program that scores how relevant that perturbation is to a trait. A single differential-expression signature or a single latent representation captures only one face of the perturbation response and can miss transcriptional variation enriched only in perturbed cells. The program library combines differential expression and co-expression, PCA and cNMF, contrastive PCA with its kernel extensions, ContrastiveVI and graph-based Hotspot modules, retaining a mean of 113.7 relatively independent programs per perturbation after redundancy removal; preserving perturbation-specific programs yielded 2.5–5.3-fold higher trait relevance scores than programs built from pooled perturbations, control cells or differential-expression summaries.

The demonstrated scope is offline analysis across 7 Perturb-seq systems, including K562, HepG2, Jurkat, TeloHAEC, THP-1, a pancreatic differentiation time course and an in vivo mouse brain screen, evaluated against 53 matched diseases and traits; the framework is currently formulated mainly for single-perturbation designs, including low-MOI pooled screens and clonally engineered perturbations.

Recovering a perturbation program depends on perturbation coverage, perturbation efficiency, the number of profiled cells and the cellular states represented in a screen; enrichment of a meta-program for disease heritability establishes genetic relevance but does not identify the causal variant or show that the perturbed gene mediates the association; power may drop for traits with sparse genetic architecture or underpowered GWAS.