MAGeCK pooled-screen analysis

When to use

Use this skill to count existing sequencing reads against a supplied guide library or compare already-counted pooled screens. Deliver the count matrix, QC, guide and gene results, contrast provenance, and a short interpretation of enrichment/depletion. This workflow analyzes screens; it does not design guides or infer gene function from a hit alone.

Runtime

The tested source installation and external-runtime caveat are in references/runtime.md. Verify both mageck --version and mageck test --help before an analysis. The bundled Python helper is standard-library only. MAGeCK itself also needs NumPy, SciPy, and the RRA binary. PDF/R reporting is optional and not needed by the helper.

The official release directory still lists 0.5.9.5 as its latest MAGeCK release. Upstream now links the separate MAGeCK2 project; these commands and the helper target MAGeCK 0.5.9.5, not an interchangeable MAGeCK2 installation. This is a local CLI workflow with no service API or authentication.

Workflow

1. Establish the library version, perturbation modality, sample names, selection direction, biological replicates, baseline material, time point, and batch. Separate sequencing lanes from independent biological replicates. A plasmid baseline and a cell day-zero baseline answer different questions. Require an explicit treatment/control contrast; the helper never silently assigns all unused samples to the control group.
2. Validate a headerless TSV library containing guide ID, DNA sequence, gene. IDs and sequences must be unique. The helper requires a count-table header beginning sgRNA, Gene, followed by unique nonnumeric sample names such as c1. MAGeCK can interpret numeric names as column indices or count values; rename them before analysis. Guide/gene IDs must have no whitespace. The helper rejects ambiguous sequences and any count/library ID or gene mismatch; resolve intentional multi-target guides explicitly upstream. These are deliberate helper restrictions; native MAGeCK also accepts other input variants.
3. For FASTQ, inspect read structure and known guide sequences to establish trimming and orientation. Use MAGeCK count, with one space-separated argument per biological sample; comma-join lanes only when they are technical replicates of that same sample. Preserve unmapped-read and count-summary evidence when mapping is poor. A zero-count guide remains in the library; do not drop it to improve QC.
4. Run QC before statistical testing. Review library representation, median reads per guide, zero fractions, Gini coefficients, and within-condition replicate correlations. The helper's 10% zero and 0.8 correlation flags are review prompts, not universal acceptance thresholds. High correlation can coexist with systematic artifacts. Read depth is not experimental cell coverage. The helper uses a raw-count population Gini; MAGeCK's native count-summary Gini uses log(count + 1) with a finite-sample correction. Do not compare their values or thresholds as the same statistic.
5. Choose normalization based on the screen. Median normalization assumes most guides are stable. For a strong global shift, supplied validated negative-control guides may support --normalization control. These must be guide IDs, one per line; a gene list is not interchangeable. Biological control samples and negative-control guides serve different roles. At least two controls must be present, and every guide assigned to a control gene must be designated a control. Supplying --control-guides also changes the RRA null distribution, even with median normalization. Record their origin and check their count distribution. MAGeCK 0.5.9.5 switches median normalization to total-count scaling for a zero median or more than 45% zero guides in any selected sample; for control normalization it evaluates the control-guide subset. Check the report's applied method, size factors, and warnings.
6. Use test for a two-group comparison. --paired requires both lists in corresponding biological order and equal length; matching lengths alone do not establish pairing. The helper reports genes at the requested FDR in both directions and retains full rankings.

…