← Về thư mục

Differential Binding Analysis - Usage Guide

Overview

Identifies differentially bound ChIP-seq regions between conditions. Distinguishes three distinct normalization problems (composition bias, trended/abundance-dependent bias, global shift) and matches each to the appropriate fix (RLE/TMM on reads-in-peaks, TMM on background bins, or spike-in scaling). Covers DiffBind (BAM + peaks), DESeq2/edgeR/PyDESeq2 (count matrix), csaw (sliding windows for global-shift-robust testing), NormR (control-aware), MAnorm2, and ChIPseqSpikeInFree.

Prerequisites

BiocManager::install(c('DiffBind', 'DESeq2', 'edgeR', 'csaw', 'apeglm', 'rtracklayer'))
BiocManager::install(c('normr', 'ChIPseqSpikeInFree'))   # optional

pip install pydeseq2 pandas numpy

Quick Start

Tell the agent what to do: - "Run DiffBind on my sample sheet with default normalization for a typical TF ChIP" - "I'm comparing H3K27ac in HDAC-inhibitor-treated vs control cells; choose the right normalization" - "Use csaw with sliding windows to detect global shifts in H3K27me3 after EZH2 inhibition" - "I have spike-in Drosophila reads; apply ChIP-Rx scaling to my DESeq2 analysis" - "Run differential binding with both reads-in-peaks and background-bin normalization and reconcile" - "Suspect global shift but no spike-in; run ChIPseqSpikeInFree for post-hoc detection" - "Run NormR for control-aware enrichment/depletion calling"

Example Prompts

Standard TF ChIP (local changes)

"Run DiffBind on 6 samples (3 ctrl, 3 treat) for FOXA1 ChIP-seq. Use the default reads-in-peaks RLE normalization since the biology is localized TF rebinding."

Global shift (HDACi / BETi / EZH2i)

"Compare H3K27me3 in EZH2-inhibitor-treated vs DMSO. EZH2 inhibition reduces H3K27me3 genome-wide; use spike-in scaling. I have Drosophila S2 read counts per sample."

Broad mark, no spike-in

"Run csaw with 1 kb windows and background-bin TMM normalization for H3K27me3 differential. Merge significant windows within 5 kb."

Suspected composition bias

"DiffBind with default settings shows almost no differential peaks but the IGV signal is obviously different. Switch to background=TRUE and re-analyze."

Post-hoc global shift detection

"I didn't run spike-in. Use ChIPseqSpikeInFree to detect whether there's a global shift in my BET-inhibitor experiment."

Sample sheet preparation

"Create a DiffBind sample sheet from my 8 BAMs + narrowPeak files. Columns: SampleID, Condition, Replicate, bamReads, bamControl, Peaks, PeakCaller."

Method reconciliation

"DiffBind reports 50 differential peaks; csaw windows reports 300. Walk through reconciliation logic and decide which to trust."

What the Agent Will Do

1. Diagnose the normalization problem: ask about expected global vs local changes (HDACi/BETi/EZH2i/dosage -> global; standard TF perturbation -> local); check MA plot loess curve after a default run
2. Choose the appropriate tool/normalization per the three-problem framework
3. Build sample sheet with SampleID, Condition, Replicate, bamReads, bamControl, Peaks, PeakCaller
4. Run differential binding:
5. DiffBind: dba.count() -> dba.normalize() -> dba.analyze()
6. DESeq2/edgeR: count matrix -> DESeqDataSet -> test
7. csaw: windowCounts() -> normFactors() -> glmQLFTest() -> merge
8. NormR: diffR() joint binomial mixture on bins
9. Spike-in integration if applicable; compute scaling factors from Drosophila/E. coli reads, pass via sizeFactors() or DiffBind spike-in fields
10. Verify normalization: dba.normalize(obj, bRetrieve = TRUE) for DiffBind; inspect MA loess curve
11. Internal-control sanity check: for spike-in, confirm blacklist regions show no signal change post-scaling
12. Output: differential peaks BED + log2FC + padj; volcano + MA + PCA + heatmap plots
13. Document: version, normalization choice, design formula, padj threshold, spike-in metadata in methods

Tips

• Always check the MA plot loess after default run. A shifted curve indicates wrong normalization. This is a 60-second sanity check that saves hours of misinterpretation.
• DiffBind 3.x defaults changed. summits=200 recenters narrow peaks (set FALSE for broad histones); blacklist filter on by default; dba.normalize() now required.
• Pre-filtering for ChIP-seq is lighter than RNA-seq. Peaks are already enriched regions; aggressive filtering removes condition-specific gains.
• LFC shrinkage is for ranking, not significance. apeglm-shrunk fold changes are better for volcano plots and downstream ranking but do NOT change padj.
• edgeR QL-F controls type-I error better than DESeq2 Wald for small replicate counts (n=2-3).
• Spike-in is the gold standard for global shifts. Apply via sizeFactors() (read-level), never multiply peak counts.
• For broad marks, use full peak width (summits=FALSE) or csaw with 1-2 kb windows. Narrow recentering loses domain-level biology.
• PCA before differential: samples should cluster by condition, not batch. If batch dominates, add batch term to design.
• Document spike-in scaling factors in methods. Patel et al 2024 found improper spike-in normalization is common; only about half of surveyed datasets had adequate matched input controls.

Troubleshooting

Almost no differential peaks despite obvious effect

Usually wrong normalization for the biology: 1. Check MA loess; if shifted off y=0, switch to background=TRUE (DiffBind) or spike-in 2. Global-shift drug? Spike-in required; no algorithmic fix works 3. Pre-filtering too aggressive -> rowSums >= 1 only

Differential peaks have wrong sign

1. Composition bias inverted the signal; switch normalization
2. Contrast direction reversed; verify contrast = c('condition', 'treat', 'ctrl') for positive log2FC = up in treat

High noise / many marginal peaks

1. Loose ENCODE-pattern input peaks; pre-filter to IDR or naive-overlap passing
2. Low FRiP / failing ChIP; see chipseq-qc; failing QC samples produce noisy differential

DiffBind very slow

dba.count(obj, bParallel = TRUE) for parallel counting. For large peak sets (>100k), consider csaw with windows instead.

Spike-in scaling gives nonsense

1. Applied to peak counts instead of sizeFactors() -> fix application layer
2. Spike-in reads not deduplicated -> re-extract from BAM
3. Spike-in saturated (too many reads) -> verify titration linearity
4. Spike-in mismatch with target -> confirm spike-in genome (Drosophila for human/mouse; E. coli for CUT&RUN/Tag)
5. Internal control changes -> see chipseq-qc and re-validate

Error: condition has only one level

Either factor has only one level or Condition column missing from sample sheet. Add at least two distinct conditions with ≥2 replicates each.

Related Skills

• chip-seq/peak-calling - Generate consensus peaks for DiffBind input
• chip-seq/chipseq-qc - Replicate concordance required before differential testing
• chip-seq/spike-in-normalization - Spike-in workflow for global-shift experiments
• chip-seq/cut-and-run-tag - E. coli spike-in (CUT&RUN) vs Drosophila (ChIP-Rx)
• chip-seq/peak-annotation - Annotate differential peaks to genes/cCREs
• differential-expression/deseq2-basics - DESeq2 patterns
• differential-expression/edger-basics - edgeR QL-F framework
• atac-seq/differential-accessibility - Parallel ATAC workflow