Bio Genome Intervals Coverage Analysis

Version Compatibility

Reference examples tested with: bedtools 2.31+, mosdepth 0.3+, samtools 1.19+, pybedtools 0.10+, numpy 1.26+.

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

samtools depth behaviour changed across versions: pre-1.13 capped depth at 8000 and truncated silently (-d 0 = unlimited); 1.13+ rewrote the subcommand with NO cap and -d/-m deprecated/ignored. Always check samtools --version before trusting a max-depth number. If code throws an error, introspect the installed tool and adapt rather than retrying.

Coverage Analysis

"Is my sequencing deep enough to answer the question?" -> Measure depth as a distribution over positions, then report median, breadth at a depth threshold, and an evenness number -- never the mean alone.

• CLI: mosdepth --by 500 prefix in.bam (windowed depth + cumulative dist), samtools coverage in.bam (per-contig depth+breadth), bedtools genomecov -ibam in.bam -bga (bedGraph track)
• Python: pybedtools.BedTool('in.bam').genome_coverage(bga=True) (pybedtools); parse prefix.mosdepth.global.dist.txt for the breadth curve

The Single Most Important Modern Insight -- Mean Depth Is a Budget, Not a Result; Report Breadth Off a Cumulative Curve

"30x WGS" describes what was paid for, not what was achieved. Coverage is a distribution over positions, and the mean is its worst summary: it is dragged up by a fat right tail (repeats, rDNA, mitochondria, PCR pileups, segmental dups) while staying blind to a hard left wall of zeros and near-zeros (GC-extreme exons, poorly-mappable regions, capture dropout). Two libraries with identical mean 30x can differ completely -- one even and callable everywhere, one spiky with 20% of the target uncallable. The mean hides both failures. Four load-bearing moves:

1. Report MEDIAN, not mean. The median is robust to the right tail. When mean/median exceeds ~1.1-1.2 the distribution is skewed and the mean is overstating typical depth -- that gap is a free evenness diagnostic.
2. Report a BREADTH / cumulative-coverage curve. "% of target >= 1x, >= 10x, >= 20x, >= 30x" is the honest summary, because adequacy is a breadth statement: a base that was not covered deeply enough is uncallable no matter how deep the rest of the genome is. mosdepth's *.mosdepth.global.dist.txt IS this curve. The killer question for any "mean = 30x" claim is "breadth at 20x?".
3. Quantify EVENNESS (CV, Fano factor, Picard fold-80, or Gini) -- an even 30x and a spiky 30x are different experiments, and a spiky library cannot be rescued by sequencing deeper (extra reads follow the same biased distribution; the holes stay holes). Fix the library (PCR-free, better capture, UMIs), not the lane count.
4. Say WHAT WAS COUNTED. A depth number is meaningless until the recipe is stated: duplicates dropped (only if MARKED first)? secondary/supplementary included? MAPQ filter? read span or fragment? mate-overlap corrected? per-base or per-region? The tools disagree on every one of these by default.

Tool Taxonomy

Decision Tree by Scenario

mosdepth -- The Modern Default

Goal: Get the median depth and the full breadth curve for a BAM in one fast pass.

Approach: Run mosdepth windowed (or whole-genome), then read the cumulative distribution file -- it already holds breadth at every depth threshold; no histogram integration needed.

mosdepth --by 500 -Q 20 sample in.bam     # --by 500 = 500 bp windows; -Q 20 = drop MAPQ<20 (repeat coverage collapses, intentionally)
# Outputs: sample.mosdepth.summary.txt (mean/min/max per chrom + total)
#          sample.mosdepth.global.dist.txt (cumulative: chrom, depth, proportion >= depth)
#          sample.regions.bed.gz (per-window mean depth)

The *.global.dist.txt rows are chrom depth proportion_of_bases_at_least_this_depth -- the breadth curve directly. Read median as the depth where proportion crosses 0.5. --fast-mode/-x is ~2x faster but SILENTLY disables mate-overlap correction -- fine for a rough WGS glance, wrong for VAF-sensitive short-insert data.

Goal: Emit a callable-region BED (NO_COVERAGE / LOW / CALLABLE / HIGH) without GATK3.

Approach: Use --quantize to bin depth and merge adjacent equal-bin runs into a compact BED.

mosdepth --quantize 0:1:4:150: callable in.bam   # bins: [0,1)=NO_COVERAGE, [1,4)=LOW, [4,150)=CALLABLE, [150,inf)=HIGH
# 4 = min callable depth (tune to caller); 150 = excessive-depth ceiling (flags rDNA/artifact pileups)
zcat callable.quantized.bed.gz | head

bedtools genomecov -- Tracks and the Histogram Default

bedtools genomecov -ibam in.bam -bga > cov.bedGraph   # -bga = bedGraph INCLUDING zero-coverage runs; -bg omits zeros
bedtools genomecov -ibam in.bam -pc -bg > frag.bedGraph # -pc = FRAGMENT coverage (mate overlap counted once); default counts reads (double-counts overlap)
bedtools genomecov -ibam in.bam -split -bg > rna.bedGraph # -split = skip N-CIGAR gaps (introns); MANDATORY for spliced RNA-seq
bedtools genomecov -ibam in.bam > hist.txt            # NO output flag = a 5-col HISTOGRAM, not a track

The bare default is a histogram, not a bedGraph -- 5 columns: chrom depth bases_at_that_depth chrom_size fraction_of_chrom, with a final genome block for the whole genome. Breadth/mean must be integrated from it yourself (sum fraction over depth >= threshold) -- which is exactly why mosdepth's ready-made dist file is preferred.

bedtools coverage -- Per-Target Stats (mind the orientation)

bedtools coverage -a targets.bed -b in.bam > per_target.bed   # stats reported FOR each A interval
bedtools coverage -a targets.bed -b in.bam -mean > mean.bed   # -mean = mean depth per A interval

As of bedtools v2.24.0 coverage is computed for the -a file (it was -b before) -- A = the regions stats are wanted for (targets), B = the reads. The default appends 4 columns to each A interval: (1) count of B features overlapping, (2) bases in A covered >=1x, (3) length of A, (4) fraction of A covered (col2/col3 = per-interval breadth). -d = per-base depth within each interval; -hist = depth histogram per interval plus an all summary; -counts = just the overlap count (faster).

samtools depth / coverage

samtools coverage in.bam                          # per-contig: rname..numreads covbases coverage(=breadth%) meandepth meanbaseq meanmapq
samtools depth -a -Q 20 -r chr1:1-100000 in.bam   # -a = report zero-depth positions; -Q = min MAPQ; -r = region
samtools depth -s in.bam                          # -s = count overlapping mate pair only ONCE (short-insert de-double-count)
samtools depth -a --reference ref.fa in.cram      # CRAM REQUIRES --reference

In samtools coverage the column literally named coverage is breadth (% bases >=1x), and meandepth is depth -- a contig with coverage=9.7 and meandepth=3.5 is 3.5x over only 9.7% of the contig (a localized pileup), NOT "9.7x coverage". samtools depth drops UNMAP/SECONDARY/QCFAIL/DUP by default (so duplicates are excluded -- but only if they were MARKED first). Without -a/-aa, zero-depth positions are omitted, so a naive sum/lines mean over-counts by dropping the zeros.

Per-Method Failure Modes

Reporting mean depth as the result

Trigger: citing "mean = 30x" as adequacy. Mechanism: mean is inflated by the repeat/rDNA tail and blind to GC/mappability holes. Symptom: a genome with large uncallable gaps looks fine. Fix: report median + breadth at the caller's threshold (mosdepth dist).