Graphs Built From Reported Results

Concise evidence-backed takeaways (claims + citations)

• Perturb‑seq enables direct linking of CRISPR perturbations to single-cell transcriptomes at scale — experimental best practices, coverage targets, capture strategies, and limitations are thoroughly reviewed (coverage recommendations, genotyping strategies, and bias sources) Perturb‑seq review (Annual Review Genetics)
• Large-scale, application-focused Perturb‑seq datasets now include genome-scale perturbations in primary human cells — 33.4M CD4+ T cells with ~26,504 gRNAs across donors/conditions provide a resource for immune regulation mapping Genome-scale CD4+ T cell perturb‑seq (2025)
• Computational methods extend Perturb‑seq: mapping & prediction — probabilistic mapping (RNA Fingerprinting), diffusion/debiasing generation (dbDiffusion), and foundation-model adapters (PertAdapt) each address different needs: assignment, unbiased fingerprinting, prediction for unseen perturbations, and transfer learning RNA Fingerprinting (probabilistic mapping) dbDiffusion (diffusion + debiasing) PertAdapt (adapter + adaptive loss)
• Convergence & functional similarity frameworks reveal shared pathways — XConTest quantifies convergent perturbations via cross-validated classifiers and module-aggregation with well-calibrated error control, aiding interpretation of overlapping regulatory programs XConTest (convergence analysis)

Limitations, blindspots, and cautions

• Coverage & genotyping: correct gRNA assignment and adequate per-gRNA cell count are required; low assignment or low knockdown efficacy reduces power and induces bias Perturb‑seq review (limitations)
• Context transfer limits: cross-cell-type or cross-species mapping raises uncertainty — RNA Fingerprinting and dbDiffusion note degraded performance when contexts diverge (e.g., p53 status differences) RNA Fingerprinting (cross-context uncertainty)
• Embedding & preprocessing dependence: methods that rely on embeddings or clustering (dbDiffusion, KeepingScore, foundation-model adapters) are sensitive to preprocessing choices (HVG selection, normalization), potentially producing batch or algorithmic biases dbDiffusion (embedding dependence)

Practical recommendations (design & analysis)

1. Plan coverage: aim for ≥100 cells per gRNA where feasible; validate gRNA efficacy and monitor assignment rates (tools: CROP-seq, direct capture strategies) Coverage & capture recommendations
2. Use probabilistic mapping for interpretation: build perturbation dictionaries and compute credible sets (RNA fingerprinting) when you need to assign mechanism or map drug/cytokine responses across contexts RNA Fingerprinting (assignments)
3. For prediction of unseen perturbations, combine data-driven embeddings with debiasing or adapter-based transfer: dbDiffusion and PertAdapt show complementary strengths — embedding quality matters; consider GO/PPI priors to constrain attention/adapters to biologically plausible edges dbDiffusion (predict unseen) PertAdapt (adapter strategy)

If you want to go further

Run a targeted BGPT analysis agent to (a) fetch specific Perturb‑seq GEO accessions, (b) compute per-perturbation pseudobulk effect sizes, (c) test embedding alignment (Rand Index) and (d) benchmark simple predictors (mean baseline, scLAMBDA, dbDiffusion) on your selected perturbations.

Selected references (open links + short extracts)

• Perturb‑seq review (Annual Review Genetics)
• RNA Fingerprinting (probabilistic mapping)
• dbDiffusion (diffusion + debiasing)
• PertAdapt (adapter for foundation models)
• Genome-scale CD4+ T cell perturb‑seq

Confidence & next steps: The visual summaries above reflect reported sample sizes, perturbation counts, and published method summaries; predictive performance statements are drawn from authors' reported PCC/benchmarks and are dataset-dependent — embedding quality and context matching critically modulate results. To reproduce/extend results on your dataset, run the AI Scientist Analysis button above to fetch raw GEO matrices and run per-perturbation pseudobulk + embedding-alignment diagnostics.