Irreproducible Discovery Rate (IDR)

IDR uses a copula mixture model to analyze the joint distribution of peak ranks across two replicate experiments. It models the data as a mixture of two components:

• A reproducible component where ranks are correlated
• A noise component where ranks are independent

The model estimates the posterior probability that each peak belongs to the reproducible component, providing a principled statistical framework rather than arbitrary overlap thresholds.

IDR produces a thresholded set of reproducible peaks by controlling the expected proportion of irreproducible discoveries. Key properties:

• An IDR threshold of 0.01 means only 1% of peaks in the final set are expected to be irreproducible
• The threshold is applied symmetrically across both replicates
• Unlike p-value cutoffs, IDR thresholds have a direct, interpretable meaning in terms of replication reliability
• The framework automatically determines the optimal signal rank cutoff

The rescue ratio quantifies how many peaks are recovered when a more lenient threshold is applied to one replicate while keeping the other fixed. This metric:

• Measures the self-consistency of the peak calling pipeline
• A high rescue ratio (>0.5) indicates that true signal is being discarded by stringent thresholds
• A low rescue ratio suggests the pipeline is already capturing most reproducible signal
• Helps diagnose whether peak calling parameters are appropriately calibrated

IDR can be applied to pseudo-replicates generated by randomly partitioning reads from a single biological sample into two pools. This technique:

• Provides a measure of technical reproducibility independent of biological variation
• Identifies the signal-to-noise ceiling of the experimental assay itself
• When combined with true biological replicates, separates technical noise from biological variability
• Essential for quality control in large-scale consortia like ENCODE

IDR evaluates reproducibility as a function of peak half-width — the width of a peak at half its maximum height. Narrower peaks typically show higher reproducibility because:

• They represent punctate, high-confidence binding events
• Broad enrichment regions often reflect non-specific signal or chromatin artifacts
• Stratifying IDR by peak width helps identify the optimal resolution for downstream analysis
• This analysis informs whether broad or narrow peak calling modes are appropriate

IDR is the gold-standard quality metric adopted by the ENCODE and modENCODE consortia for ChIP-seq peak validation. Implementation requirements:

• A minimum of two true biological replicates are required for valid IDR analysis
• Peaks passing an IDR threshold of 0.01 or 0.02 are considered reproducible
• The framework is mandatory for data submission to ENCODE portals
• IDR has been extended to ATAC-seq, DNase-seq, and ChIP-exo assays

The computational process of identifying genomic regions with statistically significant enrichment of mapped reads over background noise. IDR operates on the ranked peak lists output by peak callers, not on raw signal. The quality of input peaks—their signal-to-noise ratio and resolution—directly impacts IDR's ability to discriminate reproducible from spurious events. Common tools include MACS2 and SPP.

A quality control metric that measures the Pearson correlation between read densities on the positive and negative strands at varying shift distances. The fragment length estimated from cross-correlation is critical for IDR analysis because it determines the effective resolution of binding site detection. A high Normalized Strand Coefficient (NSC) indicates strong signal-to-noise ratio, a prerequisite for meaningful IDR thresholding.

Independent experimental replicates derived from distinct biological samples, as opposed to technical replicates from the same sample. IDR is fundamentally designed to model between-replicate agreement. True biological replicates capture the full variance of the experimental system, and IDR uses their rank concordance to distinguish reproducible biological signal from protocol-specific artifacts that may appear consistent in technical replicates.

IDR models the rank correlation of signal strength between replicate experiments rather than raw p-values or fold-change. This non-parametric approach is robust to differences in sequencing depth and signal dynamic range between replicates. The framework assumes that reproducible peaks will have consistently high ranks across replicates, while noise peaks will have uncorrelated ranks.

A statistical framework for identifying genomic loci where protein-DNA binding intensity differs significantly between experimental conditions. IDR provides the reproducible peak set that serves as the input for differential analysis. Without IDR filtering, spurious peaks inflate multiple testing burden and reduce statistical power. Tools like DESeq2 and edgeR are commonly applied to the IDR-thresholded consensus peak set.

The phenomenon where a heterozygous genetic variant causes differential transcription factor binding between maternal and paternal alleles. IDR-thresholded peaks provide a high-confidence set of binding sites for ASB analysis. Spurious peaks can generate false-positive allelic imbalances, making IDR filtering essential for accurate identification of regulatory variants that mechanistically alter binding affinity.