poisson-statistical-enrichment-testing

poisson-statistical-enrichment-testing

Summary

Applies Poisson statistics to identify genomic regions where ChIP-Seq signal significantly exceeds local background noise, computing q-value scores at single-base-pair resolution to rank peaks by statistical confidence. This is a critical intermediate step in narrow peak calling that converts pileup coverage and local bias estimates into statistical significance measures.

When to use

After extending ChIP sample reads to their predicted fragment length and constructing local lambda bias tracks (incorporating d-scaled, 1 kb, 10 kb, and genome-wide backgrounds). Apply this skill when you have aligned, deduplicated ChIP and control bedGraph pileup tracks and need to identify enriched regions by computing base-pair-level q-values before calling peak boundaries.

When NOT to use

• Input ChIP or control files have not been deduplicated — run macs3 filterdup first
• Fragment length d has not been predicted or estimated — run macs3 predictd on ChIP data
• Local bias track has not been constructed or normalized to account for multiple scales (d, slocal, llocal, genome-wide)

Inputs

• ChIP pileup bedGraph (coverage extended by predicted fragment length d)
• local lambda bedGraph (maximum bias track scaled by ChIP/control ratio)

Outputs

• q-value bedGraph track (one q-value score per base pair)

How to apply

Use macs3 bdgcmp with the -m qpois mode to compare the ChIP pileup coverage track against the scaled local lambda track (representing the expected background signal). This generates a q-value score for each genomic base pair using the Poisson model: q-values are computed as -log10(p-value) where the p-value reflects the probability of observing the ChIP count given the lambda (background) expectation. The local lambda is pre-computed by taking the maximum bias across d/2, 1 kb slocal, 10 kb llocal windows, and genome-wide backgrounds, then scaled by the ChIP-to-control sequencing depth ratio (e.g., 0.99858 in the CTCF example). The resulting bedGraph contains q-value scores suitable for thresholding (e.g., -log10(0.05) ≈ 1.301) in subsequent peak calling.