BARCS: beta-binomial regression for multivariate CRISPR screen design
Pooled CRISPR screens increasingly use longitudinal, donor-adjusted, and factorial designs, but beta-binomial screen methods have largely remained limited to pairwise comparisons. BARCS extends the library-total-conditional beta-binomial model to guide-level regression with an arbitrary design matrix, enabling direct estimation of time, covariate, and interaction effects. In four…
The BARCS (Beta-Binomial Regression for Multivariate CRISPR Screen Design) method expands upon the library-total-conditional beta-binomial model to analyze CRISPR screen data with an arbitrary design matrix. This allows for direct estimation of time, covariate, and interaction effects. In four replicate-complete Cas13 screens, incorporating an intermediate time point slightly improved the recovery of essential genes.
When the same non-targeting-control scaling rule was applied to BARCS alongside other methods, such as MAGeCK-MLE, edgeR-QL, DESeq2, and limma--voom, they all produced comparable calibration, while BARCS ranked essential genes less strongly than the alternatives. In an ordered-bin IL2RA screen, donor-adjusted BARCS identified more validated regulators with fewer total calls compared to a matched four-bin MAGeCK-MLE fit.
Additionally, cross-fitted controls revealed excess guide-level significance. Simulations demonstrated benefits from dispersion moderation and control-based denominators, but seed-specific results highlighted denominator sensitivity and substantial gene-level error due to correlated-guide aggregation rather than dispersion alone.
Aggregation-matched control scaling reduced this error but did not eliminate it. An external audit found that the reported CB2 null-discovery count vanished once full-library totals were reinstated, though this did not resolve the broader calibration concern; nominal-level calibration remained unresolved. BARCS thus offers a multivariable extension of the library-total-conditional beta-binomial model, along with an explicit account of its inference limitations.
Specifically, guide-level coefficients can be supported by independent biological libraries, while gene-level summaries and partitioned-bin designs necessitate correlation-aware aggregation or joint modeling, which the current implementation provides diagnostically rather than generatively. This distinction is emphasized due to the complexity of pooled designs, not because of any unique characteristics of the beta-binomial model.
Written by urgent.news from bioRxiv's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
