What problem does it solve? Multi-batch CRISPR screens suffer from technical variation (library lot, passage cohort, infection day, sequencing run) that can mask or mimic biological signal, and naive correction can destroy the very biology being measured. This Skill diagnoses whether batch effects dominate, selects the appropriate correction strategy, and applies it without eliminating condition signal. ## Core Features & Use Cases - Batch Diagnostics: PCA plus variance decomposition with F-statistics to quantify whether batch or condition dominates each principal component before any correction. - Four Correction Methods: ComBat empirical-Bayes with biological covariates, RUV with non-targeting control sgRNAs, SVA surrogate variables, and NTC-anchored normalization, plus the preferred alternative of modeling batch as an explicit covariate in MAGeCK MLE or Chronos design matrices. - Failure-Mode Guardrails: Detects confounded batch-condition designs, all-NaN ComBat outputs, collinear MLE design matrices, and over-correction, with post-correction validation via PCA overlap and CEGv2 PR-AUC preservation. - Use Case: You combined two screen batches and across-batch replicate Pearson is 0.78 versus 0.95 within-batch. The Skill diagnoses the batch source, applies ComBat with a condition covariate, and verifies PR-AUC is preserved before hit calling. ## Quick Start Diagnose batch effects in my multi-batch CRISPR screen with PCA and variance decomposition, then apply the appropriate correction while preserving condition signal.
