Methods¶

PerturbScape rests on a single premise: the interesting structure in a Perturb-seq experiment is the variation present in perturbed cells but absent from controls. Standard dimensionality reduction cannot isolate that - it finds whatever varies most, which is usually cell cycle, library size, or donor effects shared by both groups.

Everything follows from that premise.

flowchart LR
    A["Perturbed cells<br/>target"] --> C["Contrastive<br/>embedding"]
    B["Control cells<br/>background"] --> C
    C --> D["Structure unique<br/>to perturbation"]
    D --> E["Hotspot on the<br/>embedding graph"]
    E --> F["Local gene programs"]
    F --> G["Disease<br/>enrichment"]

Stage 1: from cells to gene programs¶

Global embeddings - what changed overall¶

A contrastive method takes two covariance structures, one from target cells and one from background cells, and finds directions that maximize target variance relative to background variance. See Contrastive Embeddings.

PerturbScape runs several of these because they disagree in useful ways, plus pca and cnmf as non-contrastive references.

Local gene programs - which genes move together¶

A global embedding gives axes of variation, not which genes act as a coordinated unit. Hotspot fills that gap: it builds a nearest-neighbour graph in the embedding space, finds genes with significant autocorrelation on that graph, and clusters them into modules. See Gene Programs and Hotspot.

Because Hotspot runs on the contrastive embedding, the modules it recovers are programs varying within perturbation-specific structure - not programs that would appear in any dataset.

Canonical contrastiveness¶

Alongside the embedding track, de-dgca computes the two classical perturbed-versus-control comparisons:

• Differential expression - which genes change in mean abundance
• Differential gene co-expression - which gene pairs change in correlation

DGCA catches something DE cannot. A perturbation can leave every gene's mean untouched while decoupling a regulatory relationship entirely. That shows up as a differential correlation and is invisible to DE.

Stage 2: from gene programs to disease¶

Programs alone do not say whether a perturbation matters for a disease. Disease Enrichment answers that in four steps: select programs carrying genetic signal for a trait, combine them into a per-perturbation meta-program, convert the top meta-program genes into SNP annotations through variant-to-gene links, and estimate a standardized heritability effect size with stratified LD score regression.

Two levels of contrast¶

The pipeline contrasts at two granularities, controlled by mode:

Disease enrichment is run per perturbation, so singular mode is what produces the perturbation-trait results published in the Data Browser.