1 Background and purpose

1.1 Why peaks are needed in genomics

In many omics assays, biological signal is not uniformly distributed across the genome. Instead, it concentrates at specific loci where a molecular event occurs, such as protein binding or an open chromatin state. Peak calling converts continuous or noisy measurements into discrete genomic intervals, enabling downstream interpretation, comparison across samples, and integration with functional genomics resources.

1.2 Inputs: aligned reads, coverage tracks, and fragments

Peak-calling workflows typically start from sequencing reads mapped to a reference genome (or to a transcriptome, depending on the assay). The signal used for peak detection may be represented as:

  • Read coverage along genomic coordinates (a base-wise or binned track)
  • Fragment counts for assays that capture fragments rather than single reads
  • Normalized signal tracks after scaling and bias correction

These input tracks summarize where sequencing evidence accumulates, which is later contrasted with an estimate of local background activity.

1.3 Definitions: peak, summit, interval, and enrichment

  • Peak: a called region where signal is enriched compared with background, often reported as an interval.
  • Summit: the coordinate within a peak interval showing the strongest signal.
  • Interval: the genomic span (start–end coordinates) assigned to a peak.
  • Enrichment: a relative measure of signal above expected background, sometimes expressed as a fold change or as a statistical effect size.

Although terminology varies across tools, these concepts remain the basis for describing and evaluating peak calls.

1.4 Common assay types using peak calling

Peak calling is widely used in sequencing-based assays such as:

  • Chromatin immunoprecipitation (ChIP-seq) for protein-DNA interactions, including transcription factors and histone modifications
  • ATAC-seq and related accessibility assays, which report open chromatin regions
  • ChIP-exo-style experiments with higher positional precision
  • CUT&RUN / CUT&Tag variants that often differ in background behavior and read distribution

Across assays, the core goal is the same: identify genomic loci with signal enriched over a comparable baseline.

2 Data preprocessing

2.1 Read processing and alignment handling

Preprocessing translates raw sequencing output into a consistent coordinate system suitable for statistical modeling. Common steps include:

  • Filtering reads by quality
  • Ensuring consistent chromosome naming and reference build
  • Handling unmapped reads and low-confidence alignments
  • Applying exclusion rules for problematic regions (tool- or project-specific)

Accurate preprocessing is essential because peak calls are sensitive to coordinate shifts, inconsistent filtering, and alignment artifacts.

2.2 Fragment-level versus read-level signal

Some assays generate signals better described at the fragment level, where each counted unit corresponds to a sequenced fragment. In paired-end protocols, fragment reconstruction can improve positional accuracy. In contrast, read-level tracks treat each read independently, which can introduce strand- or read-length-dependent biases. Tool choice often dictates whether coverage is built from reads or inferred fragments.

2.3 Normalization strategies

Normalization aims to make signal comparable across samples with differing library sizes or sequencing depth. Strategies include:

  • Scaling tracks by total mapped reads or effective counts
  • Using scaling factors derived from spike-ins or external references (when available)
  • Converting raw counts to standardized units for consistent thresholding

Normalization frequently interacts with background modeling, since over- or under-scaling can distort the apparent enrichment.

3.1 Background models and control datasets

Control datasets provide an estimate of nonspecific signal. Peak callers may use:

  • Matched input (e.g., input DNA for ChIP-seq)
  • Control IgG (for some immunoprecipitation designs)
  • Shuffled or simulated background

Background models influence both peak detection sensitivity and false positive rates. When control data quality is poor or mismatched, peak calls may become unreliable.

2.4 Bias and artifact mitigation

Sequencing data exhibits systematic biases unrelated to the underlying biology. Peak callers typically mitigate these biases to avoid calling regions enriched due to technical artifacts.

2.4.1 Mappability and GC-content considerations

Not all genomic loci are equally “mappable.” Regions with repetitive sequence may generate ambiguous alignments that inflate or deflate observed counts. Similarly, GC-content can affect sequencing efficiency and PCR amplification. Bias mitigation may involve:

  • Excluding low-mappability regions
  • Using mappability-aware normalization
  • Applying GC-aware correction models

These steps reduce the chance that peaks reflect technical variation rather than genuine enrichment.

2.4.2 Duplicate reads and PCR artifacts

PCR amplification can create duplicate molecules that over-represent certain fragments. Marking and removing duplicates helps keep counts closer to the original molecule distribution. Additionally, some pipelines treat duplicates differently depending on whether the assay is expected to produce many near-identical fragments. The appropriate approach depends on the library preparation and sequencing design.

3 Signal modeling and peak detection

3.1 Thresholding approaches

A basic strategy is to declare a peak when signal exceeds a fixed threshold—either an absolute cutoff or one defined relative to local background. Threshold-based methods are fast and simple but can struggle when noise characteristics vary across the genome or between experiments, making adaptive normalization and background estimation important.

3.2 Window-based local enrichment

Many methods operate in a sliding window. For each window, they estimate local background and identify whether signal is elevated relative to that baseline. This approach captures enrichment that is spatially localized and can adapt to broad differences in coverage patterns across genomic regions, especially when background is modeled locally rather than globally.

3.3 Statistical models for background

More sophisticated tools frame peak detection as inference: observed counts arise from a background process with additional signal where peaks exist. Such modeling improves interpretability and enables formal significance measures.

3.3.1 Count-based models and likelihood ideas

Count-based models are common because genomic signal is naturally discrete. Likelihood formulations may use distributions such as Poisson- or negative binomial-like assumptions, sometimes with overdispersion to reflect biological variability and technical noise. The core idea is to evaluate how well the data supports “background only” versus “background plus enrichment” at candidate locations.

3.4 Local maxima and summit identification

Once candidate enriched regions are identified, tools often pinpoint a summit by locating the coordinate with the highest signal within the interval. Summit detection can depend on smoothing, bin size, and fragment-shift conventions. Accurate summit placement is especially valuable for assays aiming to resolve binding sites at near-nucleotide resolution.

3.5 Peak width determination and smoothing

Peak width reflects the spatial spread of enriched signal. Width estimation can be driven by:

  • Fixed window sizes
  • Empirical smoothing of the signal track
  • Model-based extension until signal falls below a threshold

Smoothing can improve robustness against noise but may merge nearby peaks or shift summit positions if not carefully chosen.

4 Peak calling algorithms and tool families

4.1 Fixed-parameter heuristic methods

Heuristic algorithms apply rule-based criteria with limited explicit statistical inference. Examples include procedures that:

  • Use predetermined thresholds
  • Extend peak boundaries until signal drops below a cutoff
  • Merge overlapping enriched segments using simple logic

These methods can be effective when assumptions match the data, but performance may degrade when signal characteristics vary substantially.

4.2 Dynamic/statistically driven methods

Dynamic methods adjust thresholds or boundaries based on estimated local variance, background levels, or dispersion. Rather than using one global cutoff, they adapt to changing noise conditions across the genome. Such flexibility often improves robustness, particularly for samples with heterogeneous coverage.

4.3 Model-based frameworks

Model-based frameworks treat peak calling as a formal statistical task. Depending on the tool, they may:

  • Fit mixture models distinguishing background from enriched signal
  • Optimize parameters using maximum likelihood or related procedures
  • Produce significance metrics tied to inferred enrichment

This category typically provides richer confidence outputs but may require careful parameterization and adequate control data.

4.4 Graph- or signal-continuity approaches

Some approaches use concepts of signal connectivity: enriched regions form contiguous segments in a processed signal landscape. Graph-based methods can represent candidate regions as nodes connected by proximity or continuity constraints, then select segments representing peaks. These methods are useful when peak structure is better captured as a continuous signal rather than discrete counts.

4.5 Strand handling and paired-end considerations

Strand treatment is essential in assays where signal is directional or where fragments imply a shift between read starts and true binding centers. Peak callers may:

  • Separate plus and minus strand evidence
  • Reconstruct fragments using paired-end information
  • Apply strand-specific shifting to align summits

Handling paired-end data correctly often improves both localization accuracy and peak boundary placement.

5 Post-processing and refinement

5.1 Merging adjacent peaks

Initial peak detection can yield fragmented intervals for a single broader enriched locus. Merging rules combine nearby peaks when separation falls within a defined distance or when merged intervals improve coherence. The merging step affects downstream results, including peak counts, overlap statistics, and motif enrichment.

5.2 Filtering by quality metrics

Peak sets often include filters that remove dubious calls. Quality metrics can include:

  • Minimum signal strength or count thresholds
  • Consistency between replicates
  • Adequate control support (e.g., enrichment above background)

Filtering helps reduce false positives but must be balanced against the risk of discarding true but weaker peaks.

5.3 Handling broad versus narrow peaks

Different assays and targets produce peaks of varying width. Some workflows distinguish:

  • Narrow peaks (e.g., focused binding or sharp accessibility boundaries)
  • Broad peaks (e.g., certain histone marks with extended domains)

Tool parameters may switch how boundaries are defined, how smoothing is applied, and how background thresholds are assessed.

5.4 Replicate-aware adjustments

When multiple biological or technical replicates exist, peak calling can be performed per replicate followed by joint reconciliation, or performed using methods that incorporate replicate information. Replicate-aware strategies aim to keep peaks consistently observed while lowering the contribution of replicate-specific noise. The resulting peak set is often more stable for downstream comparisons.

5.5 Producing final peak sets and formats

Final outputs commonly include:

  • Genomic interval files with summit coordinates
  • Fixed-width or variable-width intervals depending on tool settings
  • Standard formats such as BED-like representations

Some pipelines also generate intermediate files (signal tracks, intermediate scores) for auditing, visualization, and reproducibility.

6 Statistical significance and confidence

6.1 P-values, false discovery rates, and multiple testing

Because peak calling evaluates many genomic locations or windows, naive p-values can produce inflated false positives. Tools therefore often report significance using multiple-testing corrections such as false discovery rate control. Understanding the meaning of corrected p-values is important for comparing peak sets across tools and experiments.

6.2 Estimating enrichment relative to background

Significance measures and effect estimates often coexist. Enrichment can be reported as fold change, normalized ratios, or model-derived statistics. While p-values reflect how surprising the observation is under a background assumption, enrichment captures magnitude, which helps interpret whether statistically significant peaks are also biologically substantial.

6.3 Reproducibility scoring and consistency across runs

Confidence can also be expressed through reproducibility metrics. For example, peaks may be scored by:

  • Overlap consistency between independent runs
  • Consistency of summit position
  • Agreement under resampling or bootstrapping

Such metrics support the practical goal of obtaining peak sets that remain stable under reasonable variations in preprocessing or sampling.

6.4 Calibration and validation of significance

Model assumptions may not perfectly match real data. Calibration checks may involve:

  • Comparing observed false positives to expectations using null datasets or permuted controls
  • Inspecting residual patterns in count models
  • Confirming that reported significance behaves sensibly across coverage regimes

Validation ensures that significance outputs are trustworthy rather than merely computed.

7 Benchmarking and evaluation

7.1 Benchmark datasets and reference standards

Evaluation typically uses curated datasets with known or high-confidence peak regions, or relies on consistent experimental designs where expected enrichment patterns are reproducible. Benchmarks may include matched controls, multi-replicate datasets, or synthetic mixtures with known ground truth.

7.2 Sensitivity, specificity, and precision-recall

Performance is commonly summarized using:

  • Sensitivity (recall): the fraction of true peaks detected
  • Specificity: how few non-peaks are included
  • Precision: how many called peaks are truly enriched

Because the class imbalance is large (most genomic positions are not peaks), precision-recall curves are often more informative than accuracy-like metrics.

7.3 Overlap metrics for peak sets

Overlap-based evaluation compares predicted peaks to reference intervals. Metrics include intersection-over-union variants, reciprocal overlap measures, and count-based matching schemes. These metrics depend on how peak boundaries are defined, making width and summit handling crucial when comparing methods.

7.4 Visual QC and diagnostic plots

Qualitative inspection remains valuable. Diagnostic plots may show:

  • Signal enrichment profiles around peaks
  • Background distributions versus signal
  • Genome-wide coverage heatmaps

Visual QC can reveal artifacts like systematic biases, overly aggressive thresholds, or failure to distinguish broad from narrow signal.

7.5 Downstream impact on analyses

Ultimately, peak calling is an upstream step. Evaluation therefore considers downstream effects such as:

  • Motif discovery stability
  • Consistency of inferred regulatory elements
  • Performance in classification or regression tasks that use peak features

A method that optimizes detection accuracy may not always maximize downstream biological utility if its outputs are less compatible with downstream modeling.

8 Parameters and practical considerations

8.1 Choosing input controls and normalization

Selecting appropriate controls and normalization settings largely determines reliability. For example, input controls should match the immunoprecipitation protocol and sequencing depth handling. Normalization should reflect both global library size differences and, when relevant, known biases in the experimental design.

8.2 Parameter tuning and sensitivity analysis

Peak calling requires parameter choices such as thresholds, window sizes, smoothing parameters, and peak merging distances. Sensitivity analysis evaluates how peak sets change when parameters vary within reasonable ranges. This helps identify settings that are robust rather than narrowly tuned to one dataset.

8.3 Read depth, sequencing depth, and saturation effects

With higher sequencing depth, the detectability of weaker peaks improves, but only up to a point. Beyond certain depths, saturation occurs: additional reads contribute diminishing new information, while technical noise may dominate. Understanding coverage-to-peak relationships can guide whether to allocate resources to deeper sequencing or improved experimental design.

8.4 Computational performance and scaling

Peak calling must scale to large genomes and high-throughput experiments. Computational performance depends on:

  • Bin sizes and window stepping
  • Complexity of statistical models
  • Replicate handling strategy

Efficient implementations, parallelization, and careful management of intermediate data can materially reduce runtime and storage requirements.

9 Outputs, interpretation, and downstream usage

9.1 Peak annotation workflows

After peak detection, peaks are mapped to genomic features such as promoters, enhancers, gene bodies, or intergenic regions. Annotation workflows often involve:

  • Defining promoter regions relative to transcription start sites
  • Using standardized gene models
  • Considering distance-based or overlap-based assignments

These choices affect which biological interpretations become possible.

9.2 Motif analysis and functional enrichment

Peaks frequently serve as input for motif analysis to identify sequence patterns associated with binding or regulatory function. Enrichment analyses may compare nucleotide composition around summits to background sequences, or use learned motifs from known databases. Outcomes depend on accurate summit placement and consistent peak widths.

9.3 Linking peaks to genes and regulatory elements

Because regulatory influence is not strictly limited to the nearest gene, mapping peaks to targets can use:

  • Proximity heuristics (e.g., nearest gene)
  • Promoter-centric approaches
  • Chromatin interaction data when available

In practice, careful interpretation is needed because peak-to-gene assignments can vary with the chosen strategy.

9.4 Differential peak analysis overview (conceptual)

Differential peak analysis aims to identify genomic regions whose enrichment changes across conditions. Conceptually, this involves comparing signal or counts for the same regions between groups, often with statistical modeling that accounts for dispersion and multiple testing. While the exact implementation varies widely, the conceptual dependence on peak definition and normalization is a central consideration.

10 Reproducibility and best practices

10.1 Versioning tools and recording parameters

Reproducibility depends on capturing the exact software versions and parameter settings used for each run. Recording the peak caller version, reference genome build, preprocessing settings, and file formats helps ensure that results can be regenerated or audited later.

10.2 Consistent preprocessing pipelines

Reproducible analysis requires uniform preprocessing across samples, including trimming, filtering, alignment settings, and fragment construction logic. Small inconsistencies can lead to shifted signal tracks and altered peak boundaries, complicating comparisons between conditions.

10.3 Reporting standards and metadata

Clear reporting typically includes:

  • Input read counts and mapping statistics
  • Details of control usage and normalization factors
  • Peak file summaries (number of peaks, typical widths, score ranges)

Standardized metadata improves interpretability and reduces ambiguity when results are shared.

10.4 Re-running peak calling reliably across batches

For large projects, pipelines must support batch processing without drifting settings. Best practices include automated workflow management, checks for reference genome consistency, validation of track scaling, and standardized QA outputs per batch. With these measures, peak calls remain comparable across time and compute environments.