Auditing causal gene networks inferred from Perturb-seq with the screens' own controls

Causal gene networks are now routinely inferred from CRISPR Perturb-seq screens, but their error rates are rarely measured. Curated references cover few regulators, and false-discovery estimates calibrated on noise cannot detect systematic artefacts. perturbaudit turns a screen's internal controls (non-targeting cells, the knockdowns themselves, independent guides, genomic positions and guide sequences) into a verdict for every edge, based on total-effect evidence in the screen. In simulations, decoy-based false-discovery estimates tracked the true false-discovery proportion under noise and confounding, but stayed near zero when a quarter of guides had seed off-targets, while the true proportion among the top 20 edges was 0.44–0.51. Across 480 simulated runs, edges that passed the audit were 80–98% true for three network methods, against 2–12% for flagged edges. We audited 40 networks from seven inference methods on ten panels drawn from public screens. Between 0% and 98% of the top-50 testable edges were not supported by the screen's own knockdown of the claimed parent. Against an all-pairs chance baseline, every genome-scale network was enriched for supported edges (1.8–9.8-fold), while seven method–panel combinations scored at or below chance. In two genome-wide CRISPRi Perturb-seq screens, distant genes with 11–12-nt promoter seed matches were knocked down 6–15% of the time, against about 0.5% expected. K562 edges that passed the audit replicated in three cell lines profiled with the same protocol more often than flagged edges (36 of 36 comparisons) and than random pairs of matched K562 effect size (31 of 36), but less often than the same number of pairs chosen by K562 effect size alone (32 of 36). On an independent platform (HCT116), every edge set transferred poorly, and audited edges did not beat the matched null. The verdicts mark which edges carry intervention evidence in the screen; for predicting other cell lines, the screen's own effect sizes are a stronger simple baseline. perturbaudit is an MIT-licensed Python package, and the seed off-target maps and audit outputs are archived on Zenodo. Preprint (version 1). Supplementary Notes 1-7, Tables S1-S7 and Figures S1-S7 are in the supplementary PDF. Code: https://github.com/mqar96/perturbaudit (release v0.2.1).

Authors

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-15
DOI
https://doi.org/10.5281/zenodo.24246654
Primary Topic
CRISPR and Genetic Engineering
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Auditing causal gene networks inferred from Perturb-seq with the screens' own controls

Muhammad Qudduse Ahmad Rao
Zenodo (CERN European Organization for Nuclear Research)
CRISPR and Genetic Engineering
preprint

Auditing causal gene networks inferred from Perturb-seq with the screens' own controls

Muhammad Qudduse Ahmad Rao
preprint en

Abstract

Causal gene networks are now routinely inferred from CRISPR Perturb-seq screens, but their error rates are rarely measured. Curated references cover few regulators, and false-discovery estimates calibrated on noise cannot detect systematic artefacts. perturbaudit turns a screen's internal controls (non-targeting cells, the knockdowns themselves, independent guides, genomic positions and guide sequences) into a verdict for every edge, based on total-effect evidence in the screen. In simulations, decoy-based false-discovery estimates tracked the true false-discovery proportion under noise and confounding, but stayed near zero when a quarter of guides had seed off-targets, while the true proportion among the top 20 edges was 0.44–0.51. Across 480 simulated runs, edges that passed the audit were 80–98% true for three network methods, against 2–12% for flagged edges. We audited 40 networks from seven inference methods on ten panels drawn from public screens. Between 0% and 98% of the top-50 testable edges were not supported by the screen's own knockdown of the claimed parent. Against an all-pairs chance baseline, every genome-scale network was enriched for supported edges (1.8–9.8-fold), while seven method–panel combinations scored at or below chance. In two genome-wide CRISPRi Perturb-seq screens, distant genes with 11–12-nt promoter seed matches were knocked down 6–15% of the time, against about 0.5% expected. K562 edges that passed the audit replicated in three cell lines profiled with the same protocol more often than flagged edges (36 of 36 comparisons) and than random pairs of matched K562 effect size (31 of 36), but less often than the same number of pairs chosen by K562 effect size alone (32 of 36). On an independent platform (HCT116), every edge set transferred poorly, and audited edges did not beat the matched null. The verdicts mark which edges carry intervention evidence in the screen; for predicting other cell lines, the screen's own effect sizes are a stronger simple baseline. perturbaudit is an MIT-licensed Python package, and the seed off-target maps and audit outputs are archived on Zenodo. Preprint (version 1). Supplementary Notes 1-7, Tables S1-S7 and Figures S1-S7 are in the supplementary PDF. Code: https://github.com/mqar96/perturbaudit (release v0.2.1).

Zenodo (CERN European Organization for Nuclear Research)
CRISPR and Genetic Engineering
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon