Peak Calling

Peak callers model the null distribution of background reads to determine if observed signal exceeds random expectation.

• Poisson models (MACS2) assume local background rates follow a Poisson distribution
• Negative binomial models (MACS3) account for overdispersion in count data
• Zero-truncated distributions handle sparse single-cell ATAC-seq data
• Hidden Markov Models (HOMER) segment the genome into enriched and background states

The choice of statistical model directly impacts the false discovery rate and the ability to detect weak but reproducible binding events.

Accurate peak calling requires estimating the average DNA fragment length from the sequencing library to shift reads toward the true binding site.

• Strand cross-correlation analysis computes the Pearson correlation between positive and negative strand read densities at varying shift distances
• The phantom peak at the read length indicates PCR artifacts
• The true peak at the fragment length reveals the predominant insertion size
• MACS2 uses this estimate to extend single-end reads to predicted fragment length before pileup computation

Poor fragment length estimation leads to split peaks or merged adjacent peaks.

Peak callers differ in how they model the expected background signal, which critically affects sensitivity in regions with variable mappability.

• Local background (MACS2): Uses a dynamic window (default 1,000–10,000 bp) around each candidate peak to estimate the local noise rate, accounting for copy number variation and chromatin state
• Global background (early methods): Assumes a uniform genome-wide background rate, which fails in repetitive or amplified regions
• Input control subtraction: Compares ChIP signal against a matched input DNA control to remove systematic biases from chromatin structure and sequencing

Local background estimation is essential for detecting peaks in heterochromatic regions with low overall signal.

Biological replicates are essential for distinguishing reproducible binding events from technical artifacts, and peak callers employ specific frameworks to assess consistency.

• Irreproducible Discovery Rate (IDR) models the rank correlation of signal strength across replicates
• IDR produces a thresholded set of peaks that are consistent across experiments
• Overlap-based approaches (bedtools intersect) require peaks to be present in multiple replicates
• Joint modeling (JAMM) simultaneously analyzes all replicates to detect condition-specific peaks

IDR has become the ENCODE consortium standard for producing high-confidence peak sets from replicate experiments.

Normalization corrects for differences in sequencing depth and library complexity between samples to enable quantitative comparison.

• Reads Per Million (RPM): Simple depth normalization that scales counts by total mapped reads
• Normalized Read Count (NRC): Scales to the smaller library size for paired comparisons
• Quantile normalization: Forces the distribution of signal values to match across samples
• Spike-in normalization: Uses exogenous chromatin from a different species as an internal reference for absolute quantification

Without proper normalization, differential binding analysis between conditions becomes unreliable due to technical rather than biological variation.

Different types of protein-DNA interactions produce characteristic peak morphologies that inform downstream analysis.

• Punctate peaks: Sharp, narrow enrichment typical of transcription factor binding at specific motifs (100–300 bp width)
• Broad domains: Wide regions of enrichment spanning kilobases, characteristic of histone modifications like H3K36me3 and H3K9me3
• Mixed regions: Some marks like H3K4me3 show both sharp peaks at promoters and broader domains
• Bimodal peaks: Paired enrichment on both strands flanking a nucleosome-free region, typical of ATAC-seq open chromatin

Peak callers like MACS2 and SICER are specifically designed for punctate or broad peak detection respectively.