Imported from huang-sh/DeepScience (packages/science/skills/biology/chip-seq/bio-chipseq-qc/SKILL.md). Install upstream with npx skills add huang-sh/DeepScience --skill bio-chipseq-qc. Copyright stays with the author.

Version Compatibility

Reference examples tested with: deepTools 3.5+, phantompeakqualtools 1.2.2+, ChIPQC 1.42+, IDR 2.0.4+, samtools 1.19+, bedtools 2.31+, pysam 0.22+, pybedtools 0.9+, MACS2 2.2.9+, MACS3 3.0.4+.

Verify versions before relying on numerical thresholds — phantompeakqualtools has known R-version compatibility issues with R ≥ 4.0 (use kundajelab fork or pin to R 3.6).

ChIP-seq Quality Control

"Should I trust this ChIP-seq experiment?" -> Validate antibody, fragmentation, enrichment, replicate concordance, library complexity, and absence of hyper-ChIPable artifacts before committing to downstream peak calling and differential analysis.

CLI: Rscript run_spp.R -c=chip.bam -out=cc.txt (NSC/RSC), plotFingerprint -b chip.bam input.bam (enrichment shape), idr --samples rep1.np rep2.np (replicate IDR)
R: ChIPQC package (Carroll & Stark; computes the full ENCODE metric battery)
Python: pysam + pybedtools for custom FRiP and library-complexity metrics

ChIP-seq fails for many independent reasons. The QC metrics below probe distinct failure modes — passing one metric does not rescue another. Antibody failure cannot be fixed by sequencing more.

The Antibody Problem is the Real Problem

Every downstream metric is conditional on antibody specificity. "ChIP-grade" on a vendor datasheet is marketing, not validation. Run the cascade:

Step	What	Why
1. Western blot	Expected MW + KO/KD negative	Confirms the antibody hits a band of the right size and loses signal in KO
2. IP-Western	Pulls down the protein	Confirms IP recovery, not just recognition
3. ChIP-qPCR	Known positive + known negative loci	First chromatin-context test; cheap
4. ChIP-seq biological replicate	Two independent biological replicates	Reproducibility check
5. KO/KD orthogonal	ChIP in KO/KD cells	Gold-standard: signal should drop to background
6. Peptide array (histones)	Epicypher SNAP-ChIP or equivalent	Tests modification-state specificity

Histone modification cross-reactivity is universal: H3K9me2 vs H3K9me3, H3K27me2 vs H3K27me3, and H3K4me1 vs H3K4me2 antibodies routinely show 10-30% cross-reactivity. Polyclonals vary lot-to-lot. CRISPR-knockout-validated lots from CST and Epicypher are the modern standard. Always record antibody catalog number + lot in methods.

Fragment-Size Distribution is a Free Diagnostic

The fragment-size distribution from a properly prepared ChIP BAM is itself a quality readout:

Distribution shape	Interpretation
Sharp peak at ~50-100 bp (sub-nucleosomal)	Direct TF binding; expected for well-fragmented TF ChIP
Sharp peak at ~150 bp + secondary at ~300 bp	Mono- + di-nucleosomal; expected for histone ChIP
Bimodal at 150 + 300, no sub-nucleosomal	Histone-only signal; in TF ChIP, suggests trapping / hyper-ChIPable
Broad continuum 100-1000 bp	Over-sonication; biology lost; cannot be rescued
No peak structure, flat	Severe over-sonication or library prep failure

# Quick diagnostic — count fragment sizes from properly-paired reads
samtools view -f 0x2 sample.bam | awk '{print $9}' | awk '$1>0' \
    | sort -n | uniq -c | awk '{print $2, $1}' > fragment_sizes.tsv

For CUT&Tag: 25-75 bp characteristic; fragments < 25 bp are Tn5 self-tagmentation noise (see cut-and-run-tag).

QC Metric Battery with ENCODE Thresholds

Metric	Tool	TF threshold	Histone threshold	Source / rationale
FRiP (Fraction of Reads in Peaks)	bedtools / pysam / featureCounts	≥ 0.01 minimum, > 0.05 ideal	≥ 0.05, > 0.20 ideal; > 0.15 for H3K4me3	Landt 2012; ENCODE flags experiments with FRiP < 1%
NSC (Normalized Strand Cross-correlation)	phantompeakqualtools	> 1.05 marginal, > 1.10 ideal	> 1.05	Landt 2012; min = 1 (no enrichment); ratio of fragment-length CC to background
RSC (Relative Strand Cross-correlation)	phantompeakqualtools	> 0.8 marginal, > 1.0 ideal	> 0.8	Landt 2012; ratio of (fragment - background) / (phantom - background)
QualityTag	phantompeakqualtools	≥ 0 acceptable, 1-2 ideal	≥ 0	Composite based on NSC/RSC; -2 to 2 scale
NRF (Non-Redundant Fraction)	unique_pos / total	> 0.8	> 0.8	ENCODE; < 0.5 severe PCR bottleneck
PBC1 (M1 / Mdistinct)	bedtools / pysam	> 0.8	> 0.8	ENCODE; fraction of singly-occupied positions
PBC2 (M1 / M2)	bedtools / pysam	> 3	> 3	ENCODE; ratio of singletons to doubletons
JS distance (plotFingerprint)	deepTools	> 0.3	> 0.05 (broad) to > 0.3 (narrow)	Distance between cumulative signal curves IP vs Input
AUC (plotFingerprint)	deepTools	< 0.6	0.6-0.9	Input = ~0.5; lower AUC = more enrichment concentrated
Synthetic JS (plotFingerprint)	deepTools	Should ≈ measured JS	—	Sanity check vs simulated null
Replicate Spearman correlation	deepTools multiBamSummary / plotCorrelation	> 0.8 (true reps)	> 0.8 (true reps), > 0.6 (broad)	Replicates should correlate more than cross-condition
Read count per replicate	samtools flagstat	≥ 20M unique mapped	20M (narrow histone), 40-60M (broad histone)	ENCODE 2012

Practical operational rule: Compute the full battery. Reject any sample failing FRiP OR antibody validation OR fragment-size sanity check, regardless of other metrics. Failing one of NSC/RSC alone with strong FRiP can sometimes be rescued for narrow-peak biology; broad histones are more forgiving on NSC.

Hyper-ChIPable Region Detection

Teytelman 2013 (PNAS): untagged GFP, no antibody, or non-existent targets all produce "binding" signal at highly-transcribed loci (rRNA, tRNA, histone gene clusters, snoRNA hosts, mtDNA, abundant housekeeping genes). ENCODE blacklist v2 (Amemiya 2019) catches repeat-driven artifacts but NOT these hyper-ChIPable transcribed regions.

Detection:

# Top 1% input signal as cell-type-specific custom blacklist
multiBigwigSummary BED-file -b input.bw -o input_signal.npz \
    --BED genes.bed --outRawCounts input_per_gene.tsv
awk 'NR > 1' input_per_gene.tsv | sort -k4,4nr | head -n $(($(wc -l < input_per_gene.tsv) / 100)) \
    > hyper_chipable.bed

# Intersect peaks against this list; flag peaks falling in hyper-ChIPable regions
bedtools intersect -a peaks.narrowPeak -b hyper_chipable.bed -u > suspicious_peaks.bed

Disprove a suspicious peak: Required for any claim at rRNA loci, tRNA clusters, HIST1/2 clusters, mitochondrial DNA:

Motif enrichment at peak (artifact has no enrichment)
KO/KD signal loss at peak (artifact persists)
Untagged-protein control ChIP shows no signal at this locus

Many "novel binding" claims at the rDNA repeat, mtDNA, and HIST1 cluster are spurious artifacts.

Computing the Battery

FRiP

total_reads=$(samtools view -c -F 260 chip.bam)
reads_in_peaks=$(bedtools intersect -a chip.bam -b peaks.narrowPeak -u | samtools view -c -)
frip=$(echo "scale=4; $reads_in_peaks / $total_reads" | bc)

NSC / RSC / fragment length (phantompeakqualtools)

Rscript run_spp.R -c=chip.bam -savp=qc/chip_cc.pdf -out=qc/chip_cc.txt
# Output columns: filename | numReads | estFragLen | corr_estFragLen |
#                 phantomPeak | corr_phantomPeak | argmin_corr | min_corr |
#                 NSC | RSC | QualityTag

Library complexity

# NRF
total=$(samtools view -c -F 260 chip.bam)
unique=$(samtools view -F 260 chip.bam | awk '{print $1, $3, $4}' | sort -u | wc -l)
nrf=$(echo "scale=4; $unique / $total" | bc)

# PBC1, PBC2 (singletons vs distinct positions vs doubletons)
samtools view -F 260 chip.bam | awk '{print $3":"$4}' | sort | uniq -c \
    | awk '{
        if ($1 == 1) m1++;
        if ($1 == 2) m2++;
        mdist++;
      } END {
        print "M1:", m1; print "M2:", m2; print "Mdistinct:", mdist;
        print "PBC1:", m1/mdist; print "PBC2:", m1/m2
      }'

deepTools plotFingerprint

plotFingerprint \
    -b chip.bam input.bam \
    --labels ChIP Input \
    -o qc/fingerprint.pdf \
    --outRawCounts qc/fingerprint_counts.tab \
    --outQualityMetrics qc/fingerprint_qc.txt
# Inspect qc/fingerprint_qc.txt: AUC, JS distance, synthetic JS, X-intercept
# Good ChIP: AUC < 0.6 (TF), JS > 0.3 (TF); Input near diagonal (AUC ~ 0.5)

Replicate Spearman correlation

multiBamSummary bins -b rep1.bam rep2.bam rep3.bam input.bam \
    --binSize 10000 -o results.npz
plotCorrelation -in results.npz --corMethod spearman \
    --whatToPlot heatmap --plotNumbers -o corr.pdf \
    --outFileCorMatrix corr_matrix.tab
# Replicates: > 0.8 (narrow), > 0.6 (broad)
# Cross-condition reps should correlate less than within-condition

ChIPQC R package

library(ChIPQC)
samples <- read.csv('samples.csv')
qc <- ChIPQC(samples, annotation = 'hg38')
ChIPQCreport(qc, reportFolder = 'ChIPQCreport')
# Generates the full ENCODE battery report per sample in one call

ChIPQC remains Bioconductor-maintained but mature; phantompeakqualtools is the canonical NSC/RSC source.

IDR and Replicate Consistency Rules

For TFs: signal-ranked IDR with Nself/Nt consistency check; see chip-seq/peak-calling for full ENCODE workflow. Key thresholds:

True replicate IDR threshold: 0.05
Pseudoreplicate IDR threshold: 0.10 (per-rep self-consistency)
Nself/Nt rule: max(N1self, N2self) / min(N1self, N2self) ≤ 2 AND max(Nt, max(Nself)) / min(Nt, min(Nself)) ≤ 2. Failing both ratios rejects the library.

For histones: naive overlap with ≥ 40% reciprocal overlap (ENCODE default; commonly misquoted as 50%). IDR is too conservative for histone signal dynamic range.

ENCODE 3 vs ENCODE 4 Thresholds (unchanged for most QC)

Metric	ENCODE 3	ENCODE 4
FRiP minimum	1%	1% (unchanged)
NSC threshold	> 1.05	> 1.05 (unchanged)
RSC threshold	> 0.8	> 0.8 (unchanged)
NRF threshold	> 0.8	> 0.8 (unchanged)
Blacklist	v1	v2 (Amemiya 2019)
Read depth (TF)	≥ 20M unique mapped	≥ 20M unchanged
Read depth (broad histone)	≥ 40M	40-60M recommended

Most QC thresholds are stable across ENCODE versions; blacklist update is the main practical change.

Per-Tool Failure Modes

phantompeakqualtools / SPP -- R version incompatibility

Trigger: Running with R ≥ 4.0.

Mechanism: spp R package has unmaintained Boost / Rcpp dependencies; some shifts produce NaN cross-correlation values.

Symptom: NSC = NaN, RSC = NaN, or fragment length = 0 in output.

Fix: Pin to R 3.6 + spp 1.16 via conda env; OR use the kundajelab/phantompeakqualtools fork (current); OR substitute deepTools plotFingerprint for enrichment QC and macs3 predictd for fragment length.

deepTools plotFingerprint -- Wrong baseline assumption for broad marks

Trigger: Interpreting JS distance with TF threshold (> 0.3) on broad histone mark.

Mechanism: Broad marks have less concentrated signal; JS distance is naturally lower (0.05-0.15 for H3K27me3) without indicating bad ChIP.

Symptom: Reports "failed JS distance" for high-quality broad-mark ChIP.

Fix: Use mark-specific thresholds: > 0.3 for TFs and sharp histones; > 0.05 for broad histones; check AUC instead (0.6-0.9 for broad; < 0.6 for TF/sharp).

FRiP -- Computed before vs after blacklist filtering

Trigger: Calling FRiP from peak file pre- vs post-blacklist.

Mechanism: Hyper-ChIPable regions inflate "reads in peaks" because most reads at those loci are artifacts.

Symptom: FRiP looks great (>15%) but most of it is rRNA / mtDNA reads.

Fix: Apply blacklist + custom hyper-ChIPable filter BEFORE computing FRiP; or report both raw and filtered FRiP.

NRF / PBC -- Computed after deduplication

Trigger: Running NRF on a MarkDuplicates-filtered BAM.

Mechanism: Library complexity metrics measure PCR redundancy; if duplicates are already removed, NRF = 1.0 by construction (uninformative).

Symptom: NRF reports 0.99-1.0; metric is meaningless.

Fix: Compute NRF / PBC1 / PBC2 on the PRE-deduplication BAM. ENCODE-compliant pipeline: filter -> MarkDuplicates (don't remove) -> compute NRF -> filter out duplicates -> call peaks.

IDR -- Wrong rank column

Trigger: Sorting narrowPeak by signalValue (column 7) for IDR.

Mechanism: MACS signalValue scales with pile-up intensity which differs between libraries of different depth; rank correlation breaks.

Symptom: IDR returns 0 reproducible peaks despite good replicate Spearman correlation.

Fix: Sort by p-value (-k8,8nr), pass --rank p.value to IDR. ENCODE convention.

ChIPQC -- Default annotation mismatch

Trigger: Using annotation = 'hg19' on hg38-aligned data.

Mechanism: ChIPQC computes feature-context enrichment from the specified annotation; mismatch silently corrupts enrichment metrics.

Symptom: Promoter / 5'UTR / 3'UTR enrichments look wrong; replicate report metrics drift.

Fix: Match annotation to the genome the BAMs were aligned to; for custom genomes pass a TxDb object explicitly.

Reconciliation: When Metrics Disagree

Pattern	Likely cause	Action
Good FRiP, bad NSC	High background but real enrichment	Acceptable for broad marks; for TFs, check phantompeakqualtools fragment length is reasonable
Good NSC, bad FRiP	Strong cross-correlation signal but few peaks pass q-value	Library shallow OR peak caller threshold too strict; try `-p 1e-2`
Good FRiP and NSC, bad replicate correlation	Real biology + replicate-specific batch effect	Check sample swap; check sequencing batch; consider PCA
Good Rep1, bad Rep2	One replicate failed	Drop Rep2 + repeat; do NOT average metrics
All metrics fail	Antibody or fragmentation failure	Re-validate antibody (KO/KD); inspect fragment-size distribution; do not proceed
FRiP excellent at rRNA/mtDNA	Hyper-ChIPable artifact dominance	Build custom blacklist; recompute

Operational rule for proceeding with downstream analysis: Require (1) antibody validated, (2) fragment-size distribution sane, (3) FRiP, NSC, RSC pass ENCODE thresholds, (4) Nself/Nt rule satisfied for TFs OR naive overlap concordance for histones, (5) hyper-ChIPable artifacts identified and either filtered or flagged.

Common Errors

Error / symptom	Cause	Solution
`Sequence chrM not found` (multi-tool)	chrM removed from BAM but kept in genome FASTA	Match chromosome naming convention; consistently include or exclude chrM
phantompeakqualtools hangs / OOM	Default tag chunking on deep libraries	Subsample to 15-25M reads (`samtools view -s`) before running
plotFingerprint blank or near-diagonal	Input control mislabeled as ChIP	Verify sample labels; AUC ~ 0.5 = Input-like signal
IDR runs but Nself ratio always > 2	One pseudoreplicate dominates due to seed	Use sufficiently different seeds (`-s 1.5` and `-s 2.5`)
ChIPQC report missing peaks	samples.csv path columns wrong	Verify bamReads / Peaks paths; ChIPQC fails silently on missing files
Replicate Spearman > 0.95	Technical (not biological) replicates	Treat as one sample; do not report as biological replicates

References

Landt SG et al 2012 Genome Res 22:1813 (ENCODE/modENCODE QC guidelines, IDR Nself rule)
Kharchenko PV et al 2008 Nat Biotechnol 26:1351 (SPP, NSC/RSC framework)
Li Q et al 2011 Ann Appl Stat 5:1752 (IDR)
Marinov GK et al 2014 G3 4:209 (large-scale ChIP-seq QC comparison)
Teytelman L et al 2013 PNAS 110:18602 (hyper-ChIPable regions)
Amemiya HM et al 2019 Sci Rep 9:9354 (ENCODE blacklist v2)
Ramírez F et al 2016 Nucleic Acids Res 44:W160 (deepTools)
Carroll TS et al 2014 Front Genet 5:75 (ChIPQC framework)
Diaz A et al 2012 Stat Appl Genet Mol Biol 11:Article 9 (CHANCE; fingerprint-style QC)
Park PJ 2009 Nat Rev Genet 10:669 (foundational review)
Rothbart SB et al 2015 Sci Rep 5:8489 (peptide-array antibody specificity)

Related Skills

chip-seq/peak-calling - Use QC metrics to decide whether to proceed with peak calling
chip-seq/cut-and-run-tag - CUT&RUN/CUT&Tag QC differs (spike-in % aligned, fragment-size signatures)
chip-seq/spike-in-normalization - QC for spike-in carryover and Drosophila read depth
chip-seq/differential-binding - Replicate concordance required before differential testing
atac-seq/atac-qc - Parallel QC for ATAC-seq (no input control, different thresholds)
alignment-files/bam-statistics - General BAM-level QC
alignment-files/duplicate-handling - MarkDuplicates before NRF computation

bio-chipseq-qc

About