Each point is one cell profiled by single-cell RNA-seq after receiving one CRISPR guide RNA from a pooled library. Cells carrying the same guide are pulled toward a target-specific position in a 3D phenotype space (a stand-in for a UMAP/PCA embedding of transcriptomic state), but only if the guide actually produced a functional edit — a fraction of cells are "escapers" that stay near the non-targeting control (NTC) cloud even though they received a targeting guide, exactly as happens with imperfect Cas9 editing efficiency.
Hit calling uses the strictly standardized mean difference (SSMD), the standard effect-size statistic for pooled screens because it accounts for both shift and spread:
SSMD_g = || mean(cell_g) − mean(cell_NTC) ||
────────────────────────────────
sqrt( var_g + var_NTC )
edited cell position = target_g · effect_size + noise
escaper cell position = target_NTC + noise (prob. = 1 − editing_efficiency)
- Cells per guide — sample size per perturbation; more cells shrink the noise term and stabilize SSMD.
- Phenotypic effect size — how strongly a true knockout shifts transcriptional state away from control.
- Editing efficiency — fraction of cells with a given guide that are truly edited; escapers dilute the measured effect.
- Technical / dropout noise — per-cell measurement noise from transcript dropout and biological variability.
- Highlight guide — dims all other clusters and reports that guide's SSMD against the NTC cloud.
In a real Perturb-seq or CROP-seq experiment, exactly this pipeline — pool cells, sequence guide + transcriptome per cell, embed, and rank targets by an SSMD-like statistic — is how thousands of gene knockouts are screened in a single assay and validated hits are followed up with individual arrayed CRISPR knockouts.