Bio Copy Number Recurrent Cnv

Version Compatibility

Reference examples tested with: GISTIC 2.0.23, R 4.3+ with CINSignatureQuantification 1.2+; Python 3.10+ with SigProfilerAssignment 0.1+ (optional, COSMIC CN signatures).

Before using code patterns, verify installed versions match. If versions differ:

• CLI: gistic2 --help (GISTIC 2.0 is a MATLAB-compiled binary; needs the MCR runtime)
• R: packageVersion('CINSignatureQuantification')
• Python: pip show SigProfilerAssignment

GISTIC 2.0 has had no substantive release since ~2017; it is effectively frozen. It runs as a compiled binary against the MATLAB Compiler Runtime — there is no R or Python package. Verify the reference (-refgene) .mat file matches the genome build.

Recurrent and Driver Copy Number Alteration

"Which copy number changes recur across my cohort, and which gene is the driver" -> A CNV in one tumor is an observation; a CNV recurring across many tumors beyond chance is evidence of selection. GISTIC2 separates recurrent driver events from passengers by modeling a background rate and scoring each locus by how often, and how strongly, it is altered. Copy-number signatures decompose the genome-wide pattern of alterations into the mutational processes that generated them.

• CLI: gistic2 — cohort-level recurrence, focal vs broad, driver localization
• R: CINSignatureQuantification (Drews 2022); Python SigProfilerAssignment (Steele 2022 COSMIC)

How GISTIC2 Works — and Its Limits

GISTIC2 scores each genomic marker with a G-score = frequency of alteration x mean amplitude, separately for amplifications and deletions. Significance (q-value) comes from permuting events along the genome under the null that all are passengers. Ziggurat deconstruction decomposes each sample's profile into the additive arm-level and focal events that produced it, so the background rate is estimated separately for broad and focal alterations — without this, ubiquitous arm-level events swamp the focal signal. A peel-off procedure removes the contribution of each significant peak before testing the next, so one strong driver does not mask its neighbors.

Two postdoc-level caveats define how GISTIC2 output must be read:

1. q-values are cohort-size dependent. Larger N manufactures more "significant" peaks. A peak list from N=50 and one from N=500 are not comparable; recurrence frequency is the portable quantity, not the q-value.
2. GISTIC2 is only as good as its input segmentation. Oversegmented seg files produce spurious narrow peaks. The seg file must also be correctly centered on diploid — a mis-centered profile (WGD genome centered on tetraploid) inverts every call before GISTIC even runs.

Decision Tree

Running GISTIC2

# Segment file: 6 columns -- sample, chrom, start, end, num_markers, seg.mean (log2).
# It MUST be diploid-centered. Pool per-sample segments (e.g. cnvkit.py export seg).
gistic2 \
    -b gistic_output/ \
    -seg cohort.seg \
    -refgene hg38.refgene.mat \
    -genegistic 1 \
    -broad 1 \
    -brlen 0.7 \
    -conf 0.99 \
    -armpeel 1 \
    -savegene 1 \
    -gcm extreme \
    -rx 0

Key flags: -brlen 0.7 sets the focal/broad cutoff at 70% of a chromosome arm; -conf 0.99 is the peak-boundary confidence — raising it above the 0.75 default yields a wider, more conservative peak with higher confidence the true driver gene lies inside it (the trade-off is more genes per peak); -armpeel 1 peels arm-level events before focal testing; -genegistic 1 runs the gene-level test; -rx 0 keeps sex chromosomes. Output amp_genes.txt / del_genes.txt and all_lesions.txt list peaks, q-values, and genes.

Copy-Number Signatures

Goal: Decompose the genome-wide copy-number pattern into mutational processes (HRD, chromothripsis, tandem duplication, ecDNA, whole-genome doubling).

Approach: Two competing 2022 frameworks exist. Steele et al (Nature 2022) defined 21 pan-cancer CN signatures from a 48-channel feature matrix, now in COSMIC; Drews et al (Nature 2022) defined 17 signatures via the CINSignatures feature set. Quantify against one framework consistently; signatures require absolute (allele-specific) copy number.

library(CINSignatureQuantification)

# segments: data frame with columns chromosome, start, end, segVal (total CN),
# sample -- absolute copy number from ASCAT/Sequenza/FACETS, NOT relative log2.
res <- quantifyCNSignatures(segments, experimentName = 'cohort',
                            method = 'drews')
activities <- getActivities(res)   # samples x signatures exposure matrix

The critical caveat (Steele 2022): three signatures had to be discarded as oversegmentation artifacts and ten were linear combinations needing manual filtering. Signatures are sensitive to the upstream caller — Steele standardizes on ASCAT (SNP6 penalty 70; WGS across the same SNP6 positions) precisely for this reason.

Failure Modes

Comparing q-values across cohorts of different size

Trigger: Stating that cohort A has "more significant" peaks than cohort B when the cohorts differ in N.

Mechanism: GISTIC q-values fall as N rises — the same recurrence frequency clears significance in a larger cohort.

Symptom: A larger cohort appears to have more drivers purely because it is larger; peak lists do not replicate.

Fix: Compare recurrence frequency (fraction of samples altered), not q-value, across cohorts. Re-run GISTIC at matched N (subsample) if a significance comparison is unavoidable.

Oversegmented input produces spurious peaks

Trigger: Feeding GISTIC a seg file from a noisy or over-fragmented segmentation.

Mechanism: GISTIC interprets every segment edge as a potential focal event boundary; fragmentation creates many narrow false peaks.

Symptom: Numerous tiny significant peaks at no known driver; peaks not replicated with a cleaner segmentation.

Fix: Quality-control the segmentation first (see copy-ratio-segmentation); merge over-fragmented segments before pooling the cohort seg file.

Mis-centered seg file inverts everything

Trigger: Pooling seg files that are not diploid-centered (e.g. WGD tumors centered on tetraploid).

Mechanism: GISTIC assumes seg.mean ~ 0 is diploid; a shifted baseline turns gains into neutral and neutral into losses before any statistics run.

Symptom: Amplification and deletion peaks swapped relative to known biology; genome-wide deletion bias.

Fix: Center each sample's seg file on its true diploid baseline (anchor with allele-specific ploidy) before pooling. Do not rely on per-sample median centering for aneuploid cohorts.