← Về thư mục
📄 / / profiles / r-and-d / raw_evidence / OpenAI4S / skills / bioskills / bio-chip-seq-differential-binding / SKILL.md

name: bio-chipseq-differential-binding description: Identifies differentially bound ChIP-seq regions between conditions using DiffBind, csaw (sliding windows), DESeq2/edgeR/PyDESeq2 on count matrices, NormR (control-aware), or MAnorm2. Distinguishes three distinct normalization problems (composition bias, trended bias, global shifts) and matches each to its appropriate fix including spike-in scaling. Use when comparing ChIP-seq binding between experimental conditions, choosing normalization for global vs local changes, integrating spike-in data, or reconciling DiffBind/DESeq2 disagreement. origin: openai4s category: bioskills/chip-seq metadata: tool_type: mixed primary_tool: DiffBind third_party: name: GPTomics/bioSkills repository: https://github.com/GPTomics/bioSkills commit: d91ed3d563019e649dc854c56ccd62551359488a license: MIT

Version Compatibility

Reference examples tested with: DiffBind 3.20+, DESeq2 1.42+, edgeR 4.0+, csaw 1.36+, PyDESeq2 0.5+, NormR 1.28+, MAnorm2 1.2+, ChIPseqSpikeInFree 1.6+.

DiffBind 3.0+ changed defaults: summits=200 (was FALSE), dba.normalize() now required, blacklist filtering on by default, full library size normalization replaces reads-in-peaks. Always run packageVersion('DiffBind') and inspect dba.normalize(obj, bRetrieve=TRUE) to confirm what was applied.

Differential ChIP-seq Binding

"Compare protein-DNA binding between experimental conditions" -> Identify regions where IP signal changes significantly, accounting for sequencing depth, composition bias, trended biases, and global shifts that confound naive normalization.

Choice of normalization matters more than choice of test statistic (RLE vs TMM vs csaw bin-TMM on the same reference reads produce nearly identical results). Choose by which of the three normalization problems applies.

The Three Distinct Normalization Problems

Problem Symptom on MA plot Cause Fix
Composition bias Loess shifts off y=0 systematically Few high-signal peaks dominate read counts; small fold changes look large or inverted TMM on background 10 kb bins (csaw / DiffBind background=TRUE); NOT reads-in-peaks
Trended bias (intensity-dependent) Loess curve sweeps from + to - across abundance Library-prep efficiency varies with fragment abundance Non-linear loess offsets (csaw normOffsets); use cautiously — can over-normalize biology
Global shift (treatment changes most peaks) Loess entirely shifted off y=0; mean log2FC ≠ 0 Drug/perturbation changes the genome-wide level of binding (HDACi, BETi, EZH2i, target KD) Spike-in scaling (ChIP-Rx); no algorithmic fix works

Why this matters: Most published ChIP-seq differential analyses default to RLE/TMM on reads-in-peaks, which assumes "most peaks are unchanged." For HDAC inhibitors, BET inhibitors, EZH2 inhibitors, or any large dosage / target-knockdown experiment, this assumption is violated. The algorithm forces the median log2FC to zero, hides the real effect, and amplifies noise around the new "zero." The result can have the wrong sign.

Diagnostic: Plot MA loess on differential results. A loess curve that sweeps abundance indicates trended bias. A uniformly shifted loess indicates a global shift. Both patterns together indicate normalization is failing in two ways simultaneously.

Algorithmic Taxonomy

Tool Treats Statistical model Strength Fails when
DiffBind 3.20+ Consensus peaks summit ± 200 bp (default) DESeq2 or edgeR backend Mature; integrated counting/normalization; spike-in support; blacklist filter on Default summits=200 recenters peaks (wrong for broad marks); DBA_NORM_LIB is conservative but misses global shifts unless background=TRUE
DESeq2 (direct) Predefined peaks NB GLM Familiar; transparent RLE on reads-in-peaks fails for global shifts
edgeR (direct) Predefined peaks NB GLM with TMM; quasi-likelihood F-test Cleaner small-sample inference; QL-F controls type-I error TMM on peak counts unstable if peaks globally shifting
csaw (Lun & Smyth 2016) Sliding windows (typically 150 bp width, 50 bp shift) edgeR QL-F Gold-standard for global shifts; bin-TMM composition bias; loess for trended biases Slower; requires BAMs not count matrix; window-merge step adds complexity
NormR (Helmuth 2016) Genomic bins Binomial mixture (background + enriched) Control-aware; identifies enrichment/depletion/background simultaneously Bin-level not peak-level; older codebase; less integration with downstream tools
MAnorm2 (Tu 2021) Peaks Hierarchical model with mean-variance trend Designed for cross-condition with replicates Less widely adopted; sparse maintenance
SpikChIP (Blanco 2021) Peaks Spike-in-aware Multi-sample spike-in comparison Niche; specific spike-in protocol assumed
SpikeFlow (2024) End-to-end Snakemake pipeline: MACS2/EPIC2/EDD peaks + DESeq2 with spike-in size factors Automated; multiple normalization options (RPM/RRPM/Rx-Input/downsampling) Inherits experimental-design errors upstream
ChIPComp (Chen 2015) Peaks Joint Poisson with input Control-aware Older; less maintained
ChIPseqSpikeInFree (Jin 2020) Peaks (post-hoc) Distribution-shape inference Detects global shift WITHOUT spike-in Post-hoc heuristic only; not definitive; sanity check

Decision Tree: Choosing Normalization

Scenario Recommended Tool / parameter
Standard TF ChIP, local changes expected, balanced gain/loss RLE on reads-in-peaks DESeq2 default; DiffBind DBA_NORM_RLE
Histone marks, broad domains, local changes TMM on background 10 kb bins csaw normFactors; DiffBind background=TRUE
HDAC / BET / EZH2 inhibitor (global change) Spike-in scaling DiffBind spikein=TRUE; SpikeFlow; manual sizeFactors() from spike reads
Dosage titration, cell-cycle synchronization Spike-in scaling Same
ChIP target knockdown / degron Spike-in or matched-input control subtraction Spike-in preferred; bamCompare log2 ratio next
CUT&RUN/CUT&Tag standard E. coli spike-in (carryover) DiffBind custom scaling; see cut-and-run-tag
No spike-in available; suspect global shift ChIPseqSpikeInFree Post-hoc distribution-shape inference
Suspected trended (abundance-dependent) bias Non-linear loess csaw normOffsets
Genome-wide enrichment/depletion analysis NormR Binomial mixture
Many conditions, large peak set edgeR QL-F or DiffBind+edgeR Better type-I control than DESeq2 Wald

Spike-In Scaling Factor Calculation

ChIP-Rx (Orlando 2014) uses Drosophila chromatin spike-in added at fixed concentration BEFORE IP. Egan 2016 adds a fixed MASS of Drosophila S2 chromatin (matched to target chromatin by the ~27:1 human:fly genome-size ratio), not a fixed cell count.

RRPM (reference-adjusted reads per million):

scale_factor_i = min(N_spike_sample) / N_spike_sample_i

Apply to read counts pre-test, OR pass as sizeFactors() to DESeq2 / normFactors() to edgeR / a numeric library-size vector to DiffBind's dba.normalize(..., library=...):

spike_in_reads <- c(120000, 145000, 110000, 95000)  # per-sample Drosophila read count
sample_names <- c('ctrl_1', 'ctrl_2', 'treat_1', 'treat_2')
scale_factors <- min(spike_in_reads) / spike_in_reads
names(scale_factors) <- sample_names

# DESeq2 with spike-in size factors
sizeFactors(dds) <- 1 / scale_factors  # DESeq2 expects inverse convention

# Or directly via DiffBind 3.x (spikein = TRUE forces library = DBA_LIBSIZE_BACKGROUND internally)
dba_obj <- dba.normalize(dba_obj, spikein = TRUE,
                          normalize = DBA_NORM_LIB)

Rx-Input variant (Fursova 2019): additionally scale by input spike-in to correct IP efficiency variation.

Internal-control sanity check: After spike-in normalization, blacklist regions and constitutive housekeeping sites (U6 promoter, rRNA processing factors that are stable) should show no signal change. If they do, the normalization is broken — common causes: - Spike-in scaling applied to peak counts instead of read counts - Spike-in reads not deduplicated before scaling - Spike-in genome not filtered for high-mapq before scaling - Spike-in saturated (always 100k+ reads); check titration linearity

Per the Patel et al 2024 Nat Biotechnol survey, improper spike-in normalization is common: of 53 datasets examined, only 27 (~51%) had adequate matched input controls across conditions.

DiffBind Workflow (BAMs + Peaks)

Goal: Run DiffBind from a sample sheet to consensus peaks, normalized counts, and tested differential binding.

Approach: Build sample sheet, count reads in consensus peaks (summit-centered for narrow marks, full peak width for broad), choose normalization based on the three-problem framework above, then test with DESeq2 or edgeR backend.

library(DiffBind)

# Sample sheet: SampleID, Condition, Replicate, bamReads, bamControl, Peaks, PeakCaller
dba_obj <- dba(sampleSheet = 'samples.csv')

# Counting: summits=250 (narrow); FALSE for broad histones (use full peak width)
dba_obj <- dba.count(dba_obj, summits = 250, minOverlap = 2, bParallel = TRUE)

# Normalization — choose per the three-problem framework
dba_obj <- dba.normalize(dba_obj,                      # default: full library size
                          method = DBA_DESEQ2,
                          normalize = DBA_NORM_LIB,
                          library = DBA_LIBSIZE_FULL)

# Background bin TMM for composition bias / broad marks
# dba_obj <- dba.normalize(dba_obj, background = TRUE)

# Spike-in scaling for global shifts
# dba_obj <- dba.normalize(dba_obj, spikein = TRUE)

dba_obj <- dba.contrast(dba_obj, design = '~ Condition')
dba_obj <- dba.analyze(dba_obj, method = DBA_DESEQ2)

# Always inspect what was actually applied
dba.normalize(dba_obj, bRetrieve = TRUE)

DiffBind 3.20+ defaults (verified via bRetrieve): - summits = 200 — narrow recentering (set FALSE for broad histones) - library = DBA_LIBSIZE_FULL — full library, conservative - normalize = DBA_NORM_LIB — library size only, not reads-in-peaks RLE - Blacklist filtering on - background = FALSE — set TRUE for composition bias

csaw Workflow (Windows-Based)

Goal: Detect differential binding from sliding windows without committing to predefined peaks; robust to composition bias via background bin TMM.

Approach: Count reads in overlapping windows, normalize on 10 kb bins (composition bias) and/or loess (trended bias), test with edgeR QL-F, then merge significant windows into regions.

library(csaw)
library(edgeR)

bam_files <- c('ctrl_1.bam', 'ctrl_2.bam', 'treat_1.bam', 'treat_2.bam')
condition <- factor(c('ctrl', 'ctrl', 'treat', 'treat'))

# Window counts: 150 bp window, 50 bp spacing (sharp marks); 1-2 kb for broad
param <- readParam(minq = 30, pe = 'both', dedup = TRUE,
                    discard = import('hg38-blacklist.v2.bed'))
windows <- windowCounts(bam_files, width = 150, ext = 200, param = param)

# Composition bias via 10 kb bins (always applied for ChIP-seq)
bg_bins <- windowCounts(bam_files, bin = TRUE, width = 10000, param = param)
windows <- normFactors(bg_bins, se.out = windows)

# Optional: trended bias via non-linear loess (use cautiously)
# windows <- normOffsets(windows, se.out = TRUE)

# edgeR QL-F test
y <- asDGEList(windows)
design <- model.matrix(~condition)
y <- estimateDisp(y, design)
fit <- glmQLFit(y, design, robust = TRUE)
results <- glmQLFTest(fit, coef = 2)

# Merge significant windows within 1 kb into regions
merged <- mergeResults(windows, results$table, tol = 1000, merge.args = list(max.width = 5000))

For broad marks, increase window width to 1-2 kb and merge tolerance to 5 kb. For TFs, 150 bp window + 50 bp spacing is standard.

DESeq2 from Count Matrix

library(DESeq2)

counts <- read.delim('counts.tsv', row.names = 1, check.names = FALSE)
coldata <- data.frame(
    condition = factor(c('ctrl', 'ctrl', 'ctrl', 'treat', 'treat', 'treat')),
    row.names = colnames(counts)
)

dds <- DESeqDataSetFromMatrix(countData = counts, colData = coldata, design = ~ condition)

# Pre-filtering for ChIP-seq is LESS aggressive than RNA-seq (peaks already enriched)
keep <- rowSums(counts(dds)) >= 1  # remove only all-zero
dds <- dds[keep, ]
dds$condition <- relevel(dds$condition, ref = 'ctrl')

dds <- DESeq(dds)
res <- results(dds, alpha = 0.05)  # match independent-filtering optimization

# Optional: LFC shrinkage for ranking/visualization only — does NOT change padj
library(apeglm)
resLFC <- lfcShrink(dds, coef = 'condition_treat_vs_ctrl', type = 'apeglm')

For spike-in normalization, set sizeFactors(dds) from scaling factors before DESeq(dds). For background-bin TMM (composition bias) use csaw to compute size factors, then transfer.

Per-Tool Failure Modes

DiffBind -- summits=200 default destroys broad marks

Trigger: Default DiffBind 3.x call with histone broad marks (H3K27me3, H3K9me3).

Mechanism: summits=200 re-centers peaks to summit ± 200 bp, throwing away most of a 10-100 kb broad domain.

Symptom: Differential count for broad marks much lower than expected; signal concentrated at narrow centers of broad regions.

Fix: dba.count(obj, summits = FALSE, ...) for broad marks; OR use full-width consensus peaks; OR switch to csaw with 1-2 kb windows.

DiffBind -- DBA_NORM_RLE reads-in-peaks default reverses global shifts

Trigger: Default (legacy DiffBind < 3.0) OR explicitly setting normalize = DBA_NORM_RLE on a global-shift experiment.

Mechanism: RLE on reads-in-peaks assumes most peaks unchanged. If 80% of peaks lose signal (e.g., EZH2 inhibitor on H3K27me3), the size factors compensate by inflating the "lost" peaks' normalized values toward control levels.

Symptom: Differential results show fewer peaks changed than visually obvious; or signs are wrong (gain reported where loss occurred).

Fix: Switch to background = TRUE (bin TMM); for definitive analysis use spikein = TRUE with ChIP-Rx.

DESeq2 -- Pre-filtering removes condition-specific peaks

Trigger: Applying RNA-seq-style filter rowSums(counts) >= 10 to ChIP-seq counts.

Mechanism: A peak present in treatment but absent in control has near-zero control counts. Aggressive filtering removes truly differential peaks.

Symptom: Significantly differential gains-of-binding peaks missing from results.

Fix: rowSums(counts) >= 1 only (remove all-zero rows); accept the loss of statistical power vs. recovering true differential peaks.

csaw -- Trended bias loess over-normalizes biology

Trigger: Applying normOffsets() (loess) when the abundance-dependent shift IS the biology.

Mechanism: Loess fits a smooth curve to the MA-plot trend; if treatment uniformly increases binding at low-signal peaks (which is biology), loess interprets it as a technical trend and removes it.

Symptom: No differential peaks detected despite obvious treatment effect.

Fix: Use bin-TMM only (composition bias); apply loess only after confirming the trend is technical (e.g., it appears in IgG-only samples).

Spike-in normalization -- Scaling factor applied to wrong layer

Trigger: Multiplying peak counts by spike-in scaling factor.

Mechanism: Spike-in factors are for read-level normalization; applied to peak counts they double-correct (peak counts already reflect mapped reads).

Symptom: Effect sizes shifted by 2-10× from expected biology; internal-control regions show artifactual signal change.

Fix: Apply spike-in via sizeFactors(dds) (DESeq2) or normFactors (edgeR) BEFORE the test; or via bamCoverage --scaleFactor for browser tracks. Never multiply peak-level counts.

IDR-passing peaks not used as differential input

Trigger: Using per-replicate MACS calls (loose -p 1e-2) as DiffBind peak input.

Mechanism: Loose ENCODE-pattern peaks include many low-confidence calls; DiffBind's minOverlap = 2 may not filter aggressively enough.

Symptom: Differential results dominated by noise at marginal peaks; high false-positive rate.

Fix: Pre-filter peak input to IDR-passing peaks (TF) or naive-overlap-passing peaks (histone) before DiffBind ingestion.

Reconciliation: When Methods Disagree

Pattern Likely cause Action
DiffBind + DESeq2 directly differ wildly Different normalization (DiffBind default = library size; DESeq2 default = RLE) Force same normalization; differences should shrink to <5%
csaw windows + DiffBind peaks disagree csaw catches sub-peak local maxima or wider regions DiffBind missed Inspect IGV; csaw windows-based often more sensitive to broad/diffuse changes
Spike-in scaled + non-scaled give opposite signs Global shift present Spike-in is correct; non-scaled is fooled by composition bias
DiffBind run twice gives different results Different summits or minOverlap settings Verify via dba.normalize(obj, bRetrieve=TRUE); pin parameters in script
Few replicates, large fold changes, low padj DESeq2 dispersion estimate unstable with n=2 Switch to edgeR QL-F or DiffBind with edgeR backend
Different fold changes in DiffBind 3.x vs 2.x Default normalization changed Match settings explicitly; document version in methods

Operational rule for publication-grade: Run on ENCODE-pattern peaks (IDR or naive overlap-passing), normalize with both reads-in-peaks AND background-bin AND spike-in if available; require concordance across at least two methods. For global-shift experiments, spike-in is mandatory.

Common Errors

Error / symptom Cause Solution
Error: dba.normalize() must be called before dba.analyze() DiffBind 3.x change; not in older docs Add dba.normalize(obj) step
DiffBind very slow on many samples Sequential counting dba.count(obj, bParallel = TRUE)
Error in DESeq() with few replicates Dispersion estimation unstable Use DESeq(dds, fitType = 'parametric', sfType = 'poscounts') or switch to edgeR
Spike-in size factors mostly NA Spike-in reads not in sample sheet or wrong BAM path Verify spike-in BAM exists; load via DiffBind spikein field
All padj = NA Independent filtering too aggressive Lower alpha in results() to match intended threshold
Volcano plot inverted (down peaks on right) Contrast direction reversed contrast = c('condition', 'treat', 'ctrl') for positive log2FC = up in treat

References