Perturb-seq and Single-Cell Screens: Linking Perturbation to Programs

A classical pooled screen compresses each perturbation into a single number — enrichment or depletion. That compression is efficient and lossy: two genes with the same viability effect may act through entirely different programs, and the screen cannot tell you.

Perturb-seq and related approaches read out single-cell transcriptomes together with the perturbation identity, so each perturbation becomes a profile rather than a rank. The gain is mechanistic resolution: modules of co-regulated genes, pathway assignments, and dose-like gradients become visible.

The cost is substantial: cell numbers, sequencing depth, and analysis complexity all rise, and many screens that would work in a selection format fail in a single-cell format for statistical rather than biological reasons.

The choice between them often comes down to whether the phenotype is transcriptomic or morphological.

Cells per perturbation. The binding constraint. A perturbation represented by 20 cells yields a noisy profile; 50–200 cells is a more typical planning range for detecting a moderate program. With 5,000 genes and 100 cells each, the arithmetic quickly exceeds practical budgets — which is why focused libraries dominate.

Depth per cell. Programs are detected as coordinated shifts across many genes, so moderate depth across more cells usually beats deep sequencing of few cells.

Guide capture efficiency. The fraction of cells with an assigned guide determines usable data. Low capture rates waste the most expensive part of the experiment.

Multiple guides per gene. As in selection screens, several guides per gene allow aggregation and reduce guide-specific artefacts.

Control cells. Non-targeting guides must be represented at a few percent to define the baseline program distribution.

• • Program specificity. Two genes with similar viability phenotypes may regulate different modules; the screen separates them.
• • Graded effects. Interference-based screens produce a continuum of knockdown, and the transcriptome responds proportionally — a dose axis that selection screens lack.
• • Cell-state dependence. Responses can be compared across clusters within one experiment, revealing state-specific dependencies.
• • Combinatorial logic. Pairs of perturbations can be compared to singles, testing genetic interaction without a viability readout.

The first is the most common and the most damaging: the replicate unit is the perturbation–cell population, or the batch, not the individual cell.

Choose single-cell when: the phenotype is transcriptional and you need to know which program changed; when genes of interest cluster into pathways you cannot separate by selection; when you need dose-response information; or when the phenotype is invisible to viability selection.

Choose a selection screen when: the phenotype is growth, survival, or a sortable marker; when you need genome-wide coverage; when budget is constrained; or when the question is "which genes matter" rather than "how do they act".

A useful hybrid: a selection screen for discovery at genome scale, followed by a focused Perturb-seq on the hits to assign programs.

The last row is what distinguishes a screen paper from a resource paper, and both are legitimate — as long as the framing matches.

Single-cell screens fail on arithmetic more often than on biology. A worked example makes the constraint clear:

• • 1,000 genes × 5 guides = 5,000 perturbations.
• • At 100 cells per perturbation, that is 500,000 cells recovered.
• • At a realistic recovery rate, the number of cells loaded is higher still.
• • Sequencing at moderate depth across 500,000 cells is a different budget class from a selection screen.

This is why published single-cell screens usually target hundreds of genes, use a focused library around a pathway, or apply a selection step before profiling. The alternative — a genome-wide library at 10 cells per perturbation — produces a dataset that cannot resolve programs, and is sometimes reported as if it could.

A practical compromise: use a selection screen for discovery at scale, then build a focused Perturb-seq library around the top few hundred hits. That sequence buys program-level resolution where it matters, without paying for it genome-wide.

EDITGENE builds the models: CRISPRi-ready lines compatible with droplet capture, validated guide delivery, reporter lines for orthogonal confirmation, and arrayed re-testing of nominated perturbations.

Tell us the program or phenotype you want to resolve. Contact EDITGENE to request a project consultation or quote.

Recoverable programs depend on cells per perturbation, depth, capture efficiency, and effect size; the planning ranges here are indicative, not universal.

Single-cell screen results describe transcriptional consequences of perturbations and remain associations until confirmed by orthogonal assays.

• • Dixit A, et al. Perturb-seq: dissecting molecular circuits with scalable single-cell RNA profiling of pooled genetic screens. Cell. 2016;167:1853–1866.
• • Replogle JM, et al. Mapping information-rich genotype–phenotype landscapes with genome-scale Perturb-seq. Cell. 2022;185:2559–2575.
• • Adamson B, et al. A multiplexed single-cell CRISPR screening platform enables systematic dissection of the unfolded protein response. Cell. 2016;167:1867–1882.
• • Feldman D, et al. Optical pooled screens in human cells. Cell. 2019;179:787–799.