Differential Binding Analysis

The foundational statistical engine for differential binding. Unlike simpler Poisson models, the negative binomial distribution accounts for overdispersion—the extra-Poisson variability inherent in ChIP-seq count data caused by biological and technical replicates.

• Models the mean-variance relationship explicitly
• Handles low-count regions without inflated false positives
• Implemented in tools like DESeq2 and edgeR (adapted for ChIP-seq)

Raw read counts are confounded by sequencing depth, IP efficiency, and chromatin input. Effective normalization is critical to avoid calling differential binding on technical artifacts.

• Library size normalization: Scales counts by total mapped reads
• Trimmed Mean of M-values (TMM): Robust to outlier high-count regions
• Spike-in controls: Exogenous chromatin added in equal amounts for absolute calibration
• Loess regression: Corrects for GC-content and mappability biases across conditions

Before differential testing, a unified set of candidate binding regions must be defined across all conditions. This consensus peak set is created by merging overlapping peaks from all samples.

• Avoids testing disjoint regions that exist in only one condition
• Reduces multiple testing burden by focusing on shared loci
• Can be generated by bedtools merge or iterative overlap procedures
• Requires careful thresholding to balance sensitivity and specificity

Testing thousands of genomic loci simultaneously inflates the false discovery rate. Rigorous correction is mandatory.

• Benjamini-Hochberg procedure: Controls the expected proportion of false positives among all rejected null hypotheses
• Bonferroni correction: Conservative family-wise error rate control, often too stringent for genomic data
• Independent Hypothesis Weighting (IHW): Increases power by leveraging covariates like mean count or peak width to weight p-values adaptively

Statistical significance alone is insufficient. Biologically meaningful differential binding requires a minimum log2 fold change between conditions.

• Prevents calling trivially small, statistically significant changes
• Typical thresholds: |log2FC| > 1 (two-fold change)
• Shrunken fold changes: Bayesian shrinkage (e.g., apeglm, ashr) stabilizes estimates for low-count regions with high dispersion
• Visualized in MA plots (log ratio vs. mean average)

Biological replicates are essential for estimating within-condition variability. The Irreproducible Discovery Rate (IDR) framework quantifies the consistency of binding events across replicates.

• Models the rank correlation of signal strength between replicate pairs
• Produces a thresholded set of reproducible peaks independent of arbitrary p-value cutoffs
• Recommended by the ENCODE Consortium as a gold-standard quality metric
• Complements differential testing by ensuring input peak sets are robust before comparison

The foundational statistical model for differential binding analysis. Unlike the Poisson distribution, the negative binomial accounts for overdispersion—the extra-Poisson variability inherent in biological count data. It models read counts using two parameters: the mean and the dispersion parameter. Tools like DESeq2 and edgeR estimate this dispersion by sharing information across genomic loci, stabilizing variance estimates when the number of biological replicates is small.

The primary assay generating data for differential binding analysis. Key design considerations include:

• Biological Replicates: A minimum of 2-3 replicates per condition is essential for estimating within-group variability.
• Input DNA Control: A sonicated, non-immunoprecipitated sample used to normalize for sequencing bias and chromatin structure.
• Spike-in Controls: Exogenous chromatin from a different species added in equal amounts to each sample, enabling absolute quantitative comparisons between conditions without being confounded by total read depth normalization.

Correcting for technical artifacts is critical before differential testing. Common methods include:

• Total Library Size: Simple scaling by total mapped reads, but fails when a few loci dominate the signal.
• Trimmed Mean of M-values (TMM): Robustly estimates scale factors by trimming extreme log-fold changes, assuming most sites are not differentially bound.
• Quantile Normalization: Forces the empirical distribution of counts to be identical across samples.
• Spike-in Calibration: Uses the signal from exogenous spike-in chromatin to derive a true scaling factor independent of the biological system.

Differential binding analysis involves performing a statistical test at every candidate peak, often numbering in the tens of thousands. Controlling false positives is mandatory. The Benjamini-Hochberg procedure controls the False Discovery Rate (FDR), the expected proportion of false positives among rejected null hypotheses. An FDR threshold of 0.05 or 0.1 is standard. This is preferred over the overly conservative Bonferroni correction, which controls the family-wise error rate.

A post-hoc quality control framework that complements differential binding analysis. IDR assesses the consistency of peak calls between pairs of biological replicates. It models the rank correlation of signal strength, producing a thresholded set of reproducible peaks. Applying IDR before differential analysis ensures that the input peak set is robust and not driven by technical artifacts, increasing the reliability of downstream differential binding results.

A specialized form of differential binding analysis that compares binding intensity between maternal and paternal alleles at heterozygous loci within a single individual. This provides a perfectly controlled internal experiment, as both alleles are exposed to the same cellular environment. ASB analysis directly links non-coding genetic variants to transcription factor binding affinity, offering functional evidence for regulatory variants identified by QTL studies.