Bio Chipseq Visualization Agent Skill

Version Compatibility

Reference examples tested with: deepTools 3.5+, pyGenomeTracks 3.9+, Gviz 1.46+, EnrichedHeatmap 1.32+, ChIPseeker 1.38+, IGV 2.17+, samtools 1.19+, bedtools 2.31+.

ChIP-seq Visualization

"Visualize ChIP-seq signal around features of interest" -> Generate normalized signal tracks (bigWig), heatmaps centered on TSS/peaks, average profile plots, and genome-browser views — with normalization that supports the biological claim (within-sample vs cross-sample vs spike-in scaled).

• CLI (production): deepTools bamCoverage -> computeMatrix -> plotHeatmap / plotProfile
• CLI (config-driven tracks): pyGenomeTracks (replaces Gviz for many use cases)
• R (publication): Gviz, EnrichedHeatmap, ChIPseeker tag heatmaps
• GUI: IGV with batch scripts for reproducible screenshots

The single most consequential choice is bigWig normalization — it determines whether visual comparison reflects biology. Get this right before generating any heatmap or browser view.

bigWig Normalization Decision Tree

| Goal | Method | When to use |

|------|--------|------------|

| Within-sample profile of a single ChIP | --normalizeUsing CPM | Standard; reads per million; comparable within one library |

| Within-sample, length-aware | --normalizeUsing BPM | TPM-analog; useful for variable-width regions; less common for ChIP-seq |

| Cross-sample with equal effective depth | --normalizeUsing RPGC --effectiveGenomeSize <N> | "1x genome coverage" — assumes equal sequencing genome-wide; ENCODE convention |

| Cross-condition with global signal change | --scaleFactor <spike_in_derived> (skip --normalizeUsing) | HDACi / BETi / EZH2i; see chip-seq/spike-in-normalization |

| ChIP vs input ratio | bamCompare --operation log2 | Visualize enrichment over input |

| ChIP vs input control-subtracted | bamCompare --operation subtract | Absolute signal above background |

| ChIP vs input SES-corrected | bamCompare --operation SES | More robust to library size; uses signal-extraction-scaling |

ENCODE convention: RPGC with read-length-matched effective genome size. For visual comparison of treatment vs control on a fold-change biology, log2 bamCompare against shared input.

Spike-in scaled tracks (the right way):

# Compute scale factor from spike-in reads (ChIP-Rx Drosophila or CUT&RUN E. coli)
SCALE=$(echo "scale=6; 1.0 / $SPIKE_IN_READS_M" | bc)  # 1 per million spike reads
bamCoverage -b chip.bam -o chip.bw --scaleFactor $SCALE --binSize 10
# DO NOT also pass --normalizeUsing; mutually exclusive

deepTools Workflow

bigWig generation

# Standard within-sample (CPM)
bamCoverage -b chip.bam -o chip.bw \
    --normalizeUsing CPM --binSize 10 \
    --extendReads 200 --numberOfProcessors 8

# Cross-sample at 1x genome coverage (ENCODE)
bamCoverage -b chip.bam -o chip.bw \
    --normalizeUsing RPGC --effectiveGenomeSize 2701495761 \
    --binSize 10 --extendReads 200

# ChIP vs Input log2 ratio (visualization of enrichment)
bamCompare -b1 chip.bam -b2 input.bam -o chip_vs_input.bw \
    --operation log2 --binSize 50 --extendReads 200 \
    --pseudocount 1 --skipZeroOverZero

Signal matrix and heatmap (reference-point: TSS / peak summit)

# Compute matrix centered on TSS
computeMatrix reference-point \
    --referencePoint TSS \
    -b 3000 -a 3000 \
    -R genes.bed \
    -S chip.bw input.bw \
    -o matrix.gz \
    --outFileSortedRegions sorted_regions.bed \
    --numberOfProcessors 8 \
    --skipZeros

# Heatmap with k-means clustering (biology emerges from clusters)
plotHeatmap -m matrix.gz \
    -o heatmap.pdf \
    --kmeans 3 \
    --colorMap RdBu_r \
    --zMin -3 --zMax 3 \
    --refPointLabel TSS \
    --heatmapHeight 12 \
    --whatToShow 'heatmap and colorbar'

# Profile plot (average signal across regions)
plotProfile -m matrix.gz \
    -o profile.pdf \
    --perGroup \
    --plotTitle 'H3K4me3 around TSS'

Scale-regions (gene-body scaled to common length)

computeMatrix scale-regions \
    -R genes.bed \
    -S chip.bw \
    -b 3000 -a 3000 \
    -m 5000 \
    -o matrix_genebody.gz \
    --numberOfProcessors 8

plotProfile -m matrix_genebody.gz -o genebody_profile.pdf --perGroup

Sample correlation

multiBamSummary bins -b sample1.bam sample2.bam sample3.bam \
    --binSize 10000 -o results.npz \
    --numberOfProcessors 8

plotCorrelation -in results.npz \
    --corMethod spearman \
    --whatToPlot heatmap \
    --plotNumbers -o correlation.pdf \
    --outFileCorMatrix correlation.tab
# Replicates should correlate > 0.8 (narrow), > 0.6 (broad)

pyGenomeTracks (Modern Browser-Style Plotting)

INI-driven, config-as-code; better than Gviz for complex layouts or pipeline integration.

# tracks.ini
[x-axis]

[chip-h3k27ac]
file = h3k27ac.bw
color = darkblue
height = 3
title = H3K27ac

[chip-h3k4me3]
file = h3k4me3.bw
color = darkred
height = 3
title = H3K4me3

[peaks-narrowpeak]
file = peaks.narrowPeak
file_type = narrow_peak
color = black
height = 0.5
title = MACS peaks

[se-bed]
file = super_enhancers.bed
color = orange
height = 0.5
title = Super-enhancers

[genes]
file = genes.gtf
color = darkgreen
prefered_name = gene_name
height = 4

pyGenomeTracks --tracks tracks.ini --region chr1:1000000-1500000 -o region.pdf

For pipeline-driven figure generation across multiple regions, pyGenomeTracks is easier to script than Gviz. For one-off publication figures with complex annotation, Gviz remains useful.

R: Gviz and EnrichedHeatmap

library(Gviz)
library(GenomicRanges)
library(TxDb.Hsapiens.UCSC.hg38.knownGene)

chr <- 'chr1'; start <- 1e6; end <- 1.1e6
itrack <- IdeogramTrack(genome = 'hg38', chromosome = chr)
gtrack <- GenomeAxisTrack()
dtrack <- DataTrack(range = 'sample.bw', genome = 'hg38',
                     type = 'histogram', name = 'ChIP', col.histogram = 'darkblue')
grtrack <- GeneRegionTrack(TxDb.Hsapiens.UCSC.hg38.knownGene,
                            genome = 'hg38', chromosome = chr, name = 'Genes')
plotTracks(list(itrack, gtrack, dtrack, grtrack), from = start, to = end, chromosome = chr)

library(EnrichedHeatmap)
library(rtracklayer)

# Normalize bigWig signal to a matrix around target sites
signal <- import('sample.bw')
tss <- promoters(txdb, upstream = 0, downstream = 1)
mat <- normalizeToMatrix(signal, tss, extend = 3000, mean_mode = 'w0', w = 50)

# Heatmap with customization
EnrichedHeatmap(mat, name = 'Signal', col = c('white', 'red'),
                top_annotation = HeatmapAnnotation(lines = anno_enriched()))

ChIPseeker Tag Heatmap (R)

library(ChIPseeker)
library(TxDb.Hsapiens.UCSC.hg38.knownGene)

peaks <- readPeakFile('peaks.narrowPeak')
promoter <- getPromoters(TxDb = TxDb.Hsapiens.UCSC.hg38.knownGene,
                          upstream = 3000, downstream = 3000)
tagMatrix <- getTagMatrix(peaks, windows = promoter)

# Tag heatmap and average profile
tagHeatmap(tagMatrix, xlim = c(-3000, 3000), color = 'red')
plotAvgProf(tagMatrix, xlim = c(-3000, 3000), conf = 0.95,
             xlab = 'Distance from TSS (bp)', ylab = 'Peak density')

IGV Batch Scripts

# IGV batch script for reproducible screenshots
cat > igv.batch << 'EOF'
new
genome hg38
load chip.bw
load peaks.bed
load super_enhancers.bed
goto chr1:1000000-1100000
snapshot region1.png
goto chr2:50000000-51000000
snapshot region2.png
exit
EOF

igv.sh -b igv.batch

Per-Tool Failure Modes

bamCoverage -- --normalizeUsing and --scaleFactor conflict

Trigger: Passing both --normalizeUsing CPM and --scaleFactor X.

Mechanism: deepTools applies scaleFactor first, then normalizes; the normalization undoes the scale.

Symptom: Spike-in scaling appears to have no effect; tracks look like CPM.

Fix: Use ONE — --scaleFactor alone for spike-in; --normalizeUsing alone otherwise. Never both.

bamCompare -- log2 with zeros produces -Inf

Trigger: bamCompare --operation log2 without pseudocount; many bins have zero reads.

Mechanism: log2(0/x) = -Inf; downstream tools (plotHeatmap) may color these as NaN or fail.

Fix: Add --pseudocount 1 to both samples; or use --skipZeroOverZero to skip bins with zero in both samples.

computeMatrix -- Stranded bigWig vs unstranded reference points

Trigger: Using stranded bigWigs (separate plus/minus) with reference-point mode on a BED without strand info.

Mechanism: computeMatrix doesn't auto-detect strand; signal is plotted in genomic-strand orientation, breaking TSS-centered plots.

Fix: Use unstranded merged bigWig OR ensure BED has strand column 6.

plotHeatmap --kmeans -- Order depends on first sample only

Trigger: Using k-means with multiple samples and expecting consistent clustering.

Mechanism: k-means clusters by signal in the first -S bigWig only; other samples are plotted in the same row order.

Fix: Order samples in -S so the most-discriminating one is first; for combined clustering across samples, use --hclust or run k-means externally on combined matrix.

Spike-in scaled bigWig -- Wrong scale factor direction

Trigger: Computing scale_factor = spike_reads / 1e6 and passing to --scaleFactor.

Mechanism: deepTools multiplies signal by scaleFactor; the INVERSE is correct (sample with fewer spike reads gets larger scale factor to compensate).

Symptom: Treatment samples appear lower than control even when biology says higher.

Fix: scale_factor = MIN(spike_reads_all_samples) / spike_reads_this_sample. Always verify against known internal-control regions (blacklist should show no signal change post-scaling).

Gviz / EnrichedHeatmap -- Memory failure on whole-genome bigWigs

Trigger: Loading a 3 GB bigWig into R as a GRanges.

Mechanism: Gviz loads the entire bigWig into memory for genome-wide views.

Fix: Use chromosome parameter to restrict; use import.bw(con, which = GRanges(...)) to subset; consider pyGenomeTracks for whole-chromosome views.

pyGenomeTracks -- INI parsing strict

Trigger: Custom INI keys not recognized; or section names with spaces.

Mechanism: pyGenomeTracks expects exact key names; case-sensitive section labels.

Fix: Run make_tracks_file --trackFiles sample.bw -o tracks.ini to generate a template; modify from there.

Reconciliation: When Visualizations Disagree

| Pattern | Likely cause | Action |

|---------|--------------|--------|

| Heatmap shows enrichment; profile plot doesn't | Signal concentrated at few regions; profile averages them out | Both correct; heatmap shows distribution, profile shows central tendency |

| Replicate heatmaps differ at peak edges | Different normalization or stranded vs unstranded bigWigs | Verify bigWig parameters identical; use same --normalizeUsing |

| Spike-in scaled tracks show opposite trend from CPM | Global shift; CPM forces median to control levels | Spike-in is correct; CPM is fooled by composition |

| ChIPseeker tag heatmap differs from deepTools heatmap | ChIPseeker uses peak density; deepTools uses signal coverage | Different metrics; pick one per analysis |

| Profile plot loose-replicate band wide | Genuine biological variability OR one replicate failed | Check per-replicate metrics (chipseq-qc); don't average across failing rep |

Common Errors

| Error / symptom | Cause | Solution |

|-----------------|-------|----------|

| bigWig has all zeros | Wrong chromosome naming (chr vs no chr) | samtools view -H bam | head to check; convert if needed |

| computeMatrix "all regions skipped" | BED chromosome naming mismatches bigWig | Match seqlevels |

| plotHeatmap colors compressed | --zMin/--zMax not set; outliers dominate | Set --zMin -3 --zMax 3 or use percentile-based |

| IGV batch hangs | exit command missing; IGV waits for input | Always end batch script with exit |

| pyGenomeTracks region out of range | Region exceeds chromosome length | Verify region from samtools view -H bam |

| Spike-in scaled track has artifact stripes | Scale factor too extreme (>10x) | Verify spike-in reads adequate (>100k); check titration |

References

• Ramírez F et al 2016 Nucleic Acids Res 44:W160 (deepTools)
• Lopez-Delisle L et al 2021 Bioinformatics 37:422 (pyGenomeTracks)
• Hahne F & Ivanek R 2016 Methods Mol Biol 1418:335 (Gviz)
• Gu Z et al 2018 Bioinformatics 34:2879 (EnrichedHeatmap)
• Yu G et al 2015 Bioinformatics 31:2382 (ChIPseeker)
• Thorvaldsdóttir H et al 2013 Brief Bioinform 14:178 (IGV)
• ENCODE 2012 quality metrics (NSC/RSC; for cross-correlation context)

Related Skills

• chip-seq/peak-calling - Peak files for heatmap reference regions
• chip-seq/chipseq-qc - QC plots (fingerprint, correlation) complement visualization
• chip-seq/spike-in-normalization - Spike-in-scaled bigWig generation
• chip-seq/differential-binding - Visualize differential peak signal
• chip-seq/super-enhancers - SE region visualization in tracks
• data-visualization/genome-tracks - General genome track patterns + IGV batch + pyGenomeTracks
• data-visualization/heatmaps-clustering - General heatmap conventions