Bedgraph handling

Version Compatibility

Reference examples tested with: deeptools 3.5+, bedtools 2.31+, ucsc-bedgraphtobigwig 445+, pyBigWig 0.3.22+.

Before using code patterns, verify installed versions match. If versions differ:

• CLI: <tool> --version then <tool> --help to confirm flags
• Python: pip show <package> then help(module.function) to check signatures

bedGraphToBigWig has a hard, under-advertised input contract: the bedGraph must be LC_COLLATE=C-sorted by chrom then start, contain non-overlapping intervals, and ship with a chrom.sizes derived from the exact assembly the reads were aligned to. deepTools effective-genome-size tables are occasionally updated between releases - re-check the installed version's table. If code throws an error, introspect the installed tool and adapt the example to match the actual API rather than retrying.

bedGraph Handling

"Make me a coverage/signal track I can compare across samples and load in a browser" -> Generate a per-bin signal track, normalize it onto a common scale (or decide a spike-in is required), then convert the text bedGraph to an indexed bigWig under the strict sort/overlap/chrom.sizes contract.

• CLI: bamCoverage -b s.bam -o s.bw --normalizeUsing RPGC --effectiveGenomeSize <N>; bedtools genomecov -ibam s.bam -bga; LC_COLLATE=C sort -k1,1 -k2,2n in.bdg | bedGraphToBigWig /dev/stdin chrom.sizes out.bw
• Python: pyBigWig.open('s.bw') to read/extract; bw.intervals(chrom, start, end) returns the bedGraph rows

The Single Most Important Modern Insight -- A Raw Coverage bedGraph Is a Library-Size Artifact, and the Wrong Normalization Is Worse Than None

Column 4 of a raw coverage bedGraph is not biology - it is sequencing depth. Two libraries of identical biology sequenced to different depths produce different heights, so any cross-sample statement ("more signal at this promoter in treatment") on un-normalized tracks is a category error. The modern path skips the text intermediate entirely: deepTools bamCoverage takes BAM -> normalized bigWig in one step, because bigWig is indexed, binary, random-access and bedGraph is flat text. Three load-bearing moves:

1. Every library-size normalization (CPM/RPKM/BPM/RPGC=1x) assumes total signal is conserved across samples. They all just rescale each library to a common total (per-million reads, or to 1x genome coverage). That model is correct when signal only redistributes locally - the usual case - and actively wrong when the perturbation changes global levels (histone-mark KD, BET-bromodomain inhibitor, global pol-II collapse). A genuine 3-fold global increase becomes, after CPM/RPGC, no change - the extra signal is spread thin and rescaled away. The model is unfalsifiable from the normalized data: forcing both libraries to the same total defines away any global difference. Library-size normalization assumes the very thing under measurement does not happen.
2. There is no computational rescue for a global change after the fact. The only fix is an external ruler decided AT THE BENCH - a spike-in of fixed foreign chromatin per cell (ChIP-Rx, Orlando 2014; defined reference epigenome, Bonhoure 2014) - scaled by the spike-in reads, not the sample reads. The wet-lab decision had to be made before sequencing; with no spike-in, the global scale is unrecoverable. The mechanics live in chip-seq/spike-in-normalization; the decision (could this perturbation change global levels?) belongs here, up front.
3. bedGraph is scratch; bigWig is the artifact. The text bedGraph is the last human-readable checkpoint - awk '$4 > 1000' to find blacklist pileups, confirm the sort/overlap invariants - before opaque binary. Inspect it, then ship bigWig. Never distribute a bedGraph as a final product: it is unindexed, so a browser reads the whole file to render any region.

Normalization Taxonomy

Method	What it assumes	When to use	When WRONG
None	nothing (raw counts)	single-sample inspection only	any cross-sample comparison - depth confounds it
CPM	total mapped reads is the right denominator; total signal conserved	depth-only normalization; quick cross-sample on a common assay	a few high-coverage bins dominate (composition skew); global change
RPKM	as CPM plus bin length matters; total signal conserved	legacy default; depth + bin-length normalized	composition skew; global change; superseded by BPM for tracks
BPM (TPM-analog)	sum over all bins fixed at 1e6; total signal conserved	composition-aware cross-sample default; robust to a few dominant bins	global change (still a conserved-total rescale)
RPGC (1x)	mean genome-wide coverage = 1x; correct effective-genome-size; total signal conserved	field-standard ChIP/ATAC browser viewing; most interpretable height	wrong effective-genome-size (linear scaling error); global change
spike-in (external)	spike-in amount is constant per cell (a ruler that does not move)	global-level change plausible or under test	nothing computational - requires a bench step before sequencing

All five library-size methods share one axiom: total signal is conserved. The decision is not which library-size method, it is whether library-size normalization is legitimate at all (see Decision Tree).

Decision Tree by Scenario

Scenario	Recommended	Why
One BAM -> browser track, local redistribution	bamCoverage --normalizeUsing RPGC --effectiveGenomeSize <N>	one-step BAM->normalized bigWig; RPGC is the interpretable ChIP/ATAC standard
Cross-sample, composition skew likely	bamCoverage --normalizeUsing BPM	bins-per-million fixes the per-bin sum; robust to dominant bins
Global-level change plausible (KD/KO of a chromatin modifier, BET inhibitor)	spike-in -> chip-seq/spike-in-normalization	library-size normalization erases the global change by construction
RNA-seq coverage track	bamCoverage --filterRNAstrand or genomecov -bga -split	-split/strand handling so spliced reads do not paint introns
ChIP/ATAC track	add --extendReads (and --centerReads for footprints)	a read is a fragment END; raw read-end coverage is double-humped and wrong
Treatment vs input from raw BAMs	bamCompare -b1 chip.bam -b2 input.bam --operation log2	normalizes depth THEN does the arithmetic
Two already-normalized bigWigs	bigwigCompare --operation log2	arithmetic only - feeding un-normalized tracks manufactures a fake change
Stack N samples into a value matrix	bedtools unionbedg -header -names ...	union interval partition; feed the matrix to R/Python for testing
Sample-relatedness QC	multiBigwigSummary bins -> plotCorrelation/plotPCA	genome-wide value matrix for correlation/PCA
Need exact per-base arithmetic (not a browser)	keep bedGraph (genomecov -bga)	bedGraph is exact text; bigWig is binned/lossy
Convert finished bedGraph -> bigWig	LC_COLLATE=C sort then bedGraphToBigWig + matched chrom.sizes	the strict contract; inspect the text first

Generate a Normalized Track with bamCoverage (the modern default)

Goal: Turn one BAM into a normalized, browser-ready bigWig in a single command.

Approach: Let bamCoverage bin, normalize, and write bigWig directly; pick the normalization from the taxonomy, supply the effective-genome-size for RPGC, extend reads for ChIP/ATAC, and exclude chrX/chrM (and any spike-in contigs) from the scale-factor calculation.

BIN_SIZE=25                  # bp; smaller = finer + noisier + bigger. Match to feature width (sharp TF/ATAC 10-25; broad marks 50-200)
EFFGENOME=2913022398         # GRCh38 non-N length (faCount); use ONLY if multimappers were kept (see Effective Genome Size)

bamCoverage -b sample.bam -o sample.bw \
  --binSize $BIN_SIZE --normalizeUsing RPGC --effectiveGenomeSize $EFFGENOME \
  --extendReads --ignoreForNormalization chrX chrM -p 8

Defaults to verify: --binSize 50, --normalizeUsing None, --scaleFactor 1.0, --extendReads off, --centerReads off. For single-end ChIP supply the fragment length (--extendReads 200); paired-end infers it. --scaleFactor with --scaleFactorsMethod None is the hook for a bench-derived spike-in factor. --outFileFormat bedgraph writes the text form when the raw numbers are needed.

Generate with bedtools genomecov (text, flexible, no normalization)

-bg collapses equal-coverage runs but omits zero-coverage regions; -bga additionally tiles zeros (use when downstream tools need explicit 0s). -split is mandatory for RNA-seq so spliced reads do not paint introns. -scale 1000000/<nreads> is a crude manual RPM; deepTools is preferred for real normalization.

bedtools genomecov -ibam sample.bam -bga -split > sample.bedgraph

Convert bedGraph -> bigWig (the silent-corruption trap)

Goal: Produce a valid bigWig from a finished bedGraph without shipping a file that loads but lies.

Approach: C-locale-sort, guarantee non-overlapping intervals, derive chrom.sizes from the exact aligned-to FASTA, inspect the text, then convert.

samtools faidx ref.fa && cut -f1,2 ref.fa.fai > chrom.sizes   # chrom.sizes from the SAME FASTA the reads aligned to
LC_COLLATE=C sort -k1,1 -k2,2n sample.bedgraph > sample.sorted.bedgraph   # C locale: locale-aware sort triggers "is not case-sensitive sorted"
bedGraphToBigWig sample.sorted.bedgraph chrom.sizes sample.bw

If concatenation/merging introduced overlaps, collapse with an explicit aggregation BEFORE converting - and note max vs mean vs sum are different signals, there is no safe default:

bedtools merge -i sample.sorted.bedgraph -d 0 -c 4 -o max > sample.nonoverlap.bedgraph

bigWig round-trips losslessly: bigWigToBedGraph sample.bw out.bedgraph (optionally -chrom=chr1 -start=1000 -end=2000).

Multi-Sample Arithmetic

Goal: Compare two tracks (treatment/input, two conditions) without letting a depth difference masquerade as biology.

Approach: From raw BAMs use bamCompare, which normalizes depth THEN applies the operation; only use bigwigCompare on bigWigs that are already on a common scale.

bamCompare -b1 chip.bam -b2 input.bam -o log2ratio.bw \
  --operation log2 --pseudocount 1 --binSize 25 --scaleFactorsMethod readCount

--operation (NOT --ratio) chooses log2/ratio/subtract/add/mean/reciprocal_ratio/first/second; default log2. --scaleFactorsMethod readCount (the default) scales by library size; --scaleFactorsMethod SES (signal-extraction scaling, Diaz 2012) instead estimates the factor from the shared background bins and is more robust than readCount for SHARP/punctate marks and TF ChIP where enrichment is a small genomic fraction; it DEGRADES for broad marks (H3K27me3/H3K9me3) where the diffuse enrichment cannot be cleanly separated from background, so use readCount (or spike-in) there. --pseudocount (default 1) prevents divide-by-zero in log2/ratio but pulls low-coverage bins toward 0 - a log2 track's apparent dynamic range is partly a pseudocount+bin-size artifact, do not read fold-changes off a browser track as measured. bigwigCompare --skipZeroOverZero drops bins that are 0 in both rather than flooding the output with log2(1)=0. Stack many samples and QC relatedness:

bedtools unionbedg -i s1.bdg s2.bdg s3.bdg -header -names s1 s2 s3 > matrix.txt   # inputs must be coordinate-sorted
multiBigwigSummary bins -b s1.bw s2.bw s3.bw -o scores.npz && plotCorrelation -in scores.npz --corMethod spearman --whatToPlot heatmap -o corr.png

Read/Extract Signal with pyBigWig

import pyBigWig

bw = pyBigWig.open('sample.bw')
mean_over_region = bw.stats('chr1', 1_000_000, 1_010_000, type='mean')[0]   # binned summary, not per-base
rows = bw.intervals('chr1', 1_000_000, 1_010_000)   # the underlying bedGraph rows: (start, end, value)
bw.close()

bw.stats()/bw.values() return what the bin resolution preserved, not a faithful per-base record - coarse bins silently change the values read back.

Effective Genome Size (the two-table trap)

--effectiveGenomeSize feeds the RPGC scale factor and depends on the read-filtering regime. deepTools ships two tables that answer different questions:

Build	Non-N length (faCount; multimappers KEPT)
GRCh38	2,913,022,398
GRCh37	2,864,785,220
GRCm38 (mm10)	2,652,783,500
dm6	142,573,017
WBcel235 (C. elegans)	100,286,401

When reads were instead filtered to unique alignments / a MAPQ filter applied (the common ChIP/ATAC case), use the read-length-dependent unique-k-mer value: GRCh38 is 2,701,495,711 (50 bp), 2,805,636,231 (100 bp), 2,862,010,428 (150 bp). The two GRCh38 numbers differ ~7% at short read length. RPGC scales linearly in this value, so the error cancels for within-study ratios but surfaces as a spurious constant fold-difference on cross-study integration (a public track, a collaborator's bigWig, a track made last year at a different read length). For non-model organisms there is no table - estimate it (faCount for non-N length, or unique-k-mers on the assembly).

Per-Method Failure Modes

Comparing un-normalized tracks across samples

Trigger: browser-comparing or quantifying raw coverage bedGraphs/bigWigs. Mechanism: column 4 scales with library size. Symptom: the deeper library looks like it has "more signal" everywhere. Fix: normalize during bamCoverage; never compare --normalizeUsing None tracks.

Conserved-total assumption under a global change

Trigger: CPM/RPKM/BPM/RPGC on a perturbation that shifts global levels (chromatin-modifier KD/KO, BET inhibitor). Mechanism: every library-size method forces total signal to a constant. Symptom: a real global increase reads as no change; tracks look identical. Fix: spike-in decided at the bench -> chip-seq/spike-in-normalization. No computational rescue exists.

Unsorted / overlapping input -> corrupt bigWig

Trigger: bedGraphToBigWig on non-C-sorted or overlapping input, or chrom.sizes from the wrong assembly. Mechanism: the contract is enforced inconsistently - some violations error, others build a bigWig that loads and shows wrong heights or silently drops chromosomes. Symptom: is not case-sensitive sorted, overlapping regions, end coordinate bigger than, or a silently wrong/incomplete track. Fix: LC_COLLATE=C sort; bedtools merge -c 4 -o max/mean/sum; chrom.sizes from the exact aligned-to FASTA; harmonize chr1 vs 1.

Effective-genome-size drift

Trigger: grabbing the round 2.9e9 GRCh38 value regardless of multimapper filtering, or reusing a value across read lengths/assemblies. Mechanism: RPGC scales linearly in the value; the two tables differ ~7%. Symptom: invisible within a study; a spurious constant fold-difference on cross-study integration. Fix: match the value to the read length AND filtering regime; estimate it for non-model organisms.

Bin-size aliasing

Trigger: a bin wider than ~half the feature, or comparing tracks built at different binSizes. Mechanism: binSize is a low-pass filter chosen once; a feature narrower than ~2 bins is averaged down or straddles a boundary (a phase artifact - replicates disagree by bin alignment). Symptom: sharp peaks shrink or split; bin-for-bin ratios meaningless at boundaries. Fix: match binSize to feature width (sharp TF/ATAC 10-25 bp, broad marks 50-200 bp); compared tracks MUST share binSize. --smoothLength is cosmetic, it cannot recover discarded information.

ChIP/ATAC track without extendReads

Trigger: bamCoverage/genomecov on ChIP/ATAC without --extendReads. Mechanism: a read marks a fragment END, not the fragment. Symptom: double-humped peaks with a central dip; biased boundaries and quantification. Fix: --extendReads (paired-end infers; single-end supply the fragment length). RNA-seq mirror trap: without -split spliced reads paint introns.

Quantitative Thresholds

Threshold	Source	Rationale
binSize default 50 bp; sharp TF/ATAC 10-25 bp, broad marks 50-200 bp	deepTools default + feature-width matching	binSize is a low-pass filter; finer is noisier/bigger, coarser aliases sharp features
GRCh38 effGenome 2,913,022,398 (multimappers kept)	deepTools faCount table	non-N genome length for the RPGC denominator
GRCh38 effGenome 2.70-2.86e9 by read length (unique alignments)	deepTools unique-k-mer table	~7% below the non-N value; use when MAPQ/uniqueness-filtered
pseudocount default 1 (log2/ratio)	deepTools default	prevents divide-by-zero; biases low-coverage bins toward 0
single-end fragment length ~200 bp (--extendReads 200)	typical sonicated ChIP fragment	wrong value distorts peak width; paired-end infers it
compared tracks must share binSize	signal-processing constraint	different grids make bin-for-bin ratios meaningless

Common Errors

Error / symptom	Cause	Solution
is not case-sensitive sorted	locale-aware sort	LC_COLLATE=C sort -k1,1 -k2,2n (works on a login node, fails in the scheduler when $LC_* differ)
overlapping regions in bedGraph file	concatenated/merged tracks	bedtools merge -c 4 -o max/mean/sum (choose the aggregation deliberately)
end coordinate N bigger than ...	chrom.sizes from a different assembly/patch	derive chrom.sizes from the exact aligned-to FASTA (samtools faidx + cut -f1,2)
Whole chromosomes missing from the bigWig, no error	chr1 vs 1 / MT vs chrM naming mismatch	harmonize naming across bedGraph and chrom.sizes
Track line breaks sort/conversion	track type=bedGraph ... header row	remove the track line before sort/bedGraphToBigWig
RPGC normalization fails	--effectiveGenomeSize not supplied	pass the correct value for the build, read length, and filtering
Spliced reads paint introns	no -split (genomecov) / wrong RNA mode	genomecov -bga -split or bamCoverage --filterRNAstrand

References

• Ramírez F, Ryan DP, Grüning B, et al. 2016. deepTools2: a next generation web server for deep-sequencing data analysis. Nucleic Acids Res 44:W160-W165.
• Kent WJ, Zweig AS, Barber G, Hinrichs AS, Karolchik D. 2010. BigWig and BigBed: enabling browsing of large distributed datasets. Bioinformatics 26:2204-2207.
• Quinlan AR, Hall IM. 2010. BEDTools: a flexible suite of utilities for comparing genomic features. Bioinformatics 26:841-842.
• Orlando DA, Chen MW, Brown VE, et al. 2014. Quantitative ChIP-Seq normalization reveals global modulation of the epigenome. Cell Reports 9:1163-1170.
• Bonhoure N, Bounova G, Bernasconi D, et al. 2014. Quantifying ChIP-seq data: a spiking method providing an internal reference for sample-to-sample normalization. Genome Res 24:1157-1168.
• Diaz A, Park K, Lim DA, Song JS. 2012. Normalization, bias correction, and peak calling for ChIP-seq. Stat Appl Genet Mol Biol 11:Article 9.

Related Skills

• coverage-analysis - Per-base depth generation and distribution-vs-mean diagnostics feeding bedGraph tracks
• bigwig-tracks - Reading, extracting, and writing the bigWig deliverable this skill produces
• chip-seq/spike-in-normalization - The bench-decided external-reference scaling when a global change makes library-size normalization wrong
• chip-seq/chipseq-visualization - Render the normalized signal tracks built here
• atac-seq/footprinting - Consumes high-resolution coverage/bigWig signal over motif sites
• data-visualization/genome-tracks - Render the bedGraph/bigWig tracks for figures