Coverage / Read Depth

Coverage, or read depth, is how many sequencing reads align across a given base of the reference genome. “30X whole-genome sequencing” means the average base is covered by about 30 reads. The average is a budgeting number. What determines whether you can call a variant at position chr7:117,559,590 is the depth at that position.

How it works

The textbook estimate is the Lander-Waterman expression: mean depth = LN/G, where L is read length, N is number of reads, and G is genome size ¹. For a 2×150 bp run producing 400 million read pairs against a 3.1 Gb genome: (300 × 400e6) / 3.1e9 ≈ 39X raw. That is raw depth, before alignment. After you discard unmapped reads, duplicates, and reads with MAPQ 0 in segmental duplications, the usable depth an aligner-aware caller sees typically lands 10-20% lower.

Depth is not uniform. Under an idealized Poisson model with mean 30, the chance a base gets fewer than 10 reads is small, but real data has extra-Poisson variance driven by GC content, PCR amplification bias, and mappability ¹. Regions at the GC extremes (promoters, first exons of many genes) and regions with near-identical paralogs are systematically underserved. Roughly 1-3% of a nominal 30X PCR-free genome ends up below 10X, and a few megabases sit at zero no matter how deep you go.

Why 30X is the conventional target for germline WGS: it is where single-sample heterozygous SNV sensitivity plateaus. A het site expects half its reads to carry the alternate allele. At 10X, a binomial draw gives 0 or 1 alt reads about 1% of the time, and callers reject those. Genotyping error rates fall sharply with coverage and with base quality filtering, and low-coverage data skews specifically toward calling heterozygotes as homozygous reference ². Above ~30X the sensitivity gain per additional read is small for SNVs, though indels, copy-number segmentation, and mosaic or somatic detection keep improving. Copy-number methods work on binned depth and need enough reads per bin for the segmentation to be stable ³.

Exome is the other regime. Capture concentrates reads on ~1-2% of the genome, so 100X mean exome depth costs far less than 100X genome. The tradeoff is uneven capture efficiency and poorer coverage of GC-rich first exons; genome sequencing covers coding regions more completely and evenly than exome at matched cost points ⁴. Hybrid designs (high-depth exome plus low-depth genome) exist precisely to buy both coding sensitivity and genome-wide structural signal ⁵.

In your own data

Your depth lives in the CRAM/BAM. Compute it with mosdepth, which is fast and gives you the distribution rather than a single number ²:

mosdepth --by 1000 --fast-mode --mapq 20 \
  --threads 4 sample sample.cram

That writes sample.mosdepth.global.dist.txt (cumulative fraction of the genome at ≥ each depth), sample.mosdepth.summary.txt (mean per chromosome), and sample.regions.bed.gz (per-1kb depth). The number to care about from the global dist file is the fraction of the genome at ≥10X and ≥20X. For a good PCR-free 30X genome, expect roughly 95%+ at ≥10X. If you are checking a specific gene, feed a BED instead of --by 1000 and read the per-exon values.

In a VCF, per-site depth is the DP field in FORMAT, and allelic depth is AD (ref,alt counts). Before you believe any single variant, look at both. A het call with AD of 28,3 is probably a mapping or sequencing artifact. DP of 250 at a nominal 30X genome means a collapsed repeat, and calls there are unreliable. DP of 4 means you have no real information. GATK’s default hard filter thresholds will not catch all of these for you.

Three mistakes we see constantly. First, comparing a vendor’s “raw coverage” to another’s “mean mapped deduplicated coverage” as if they are the same metric. Ask which one is quoted. Second, trusting a ./. or 0/0 call in a low-depth region as evidence of absence; unless you have a gVCF with reference confidence blocks, absent is not the same as reference. Third, ignoring chrX and chrY depth, which should be about half autosomal in an XY sample and will immediately flag sample swaps or contamination.

Limitations

Depth says nothing about whether reads are placed correctly. Ten reads with MAPQ 0 in a paralogous region are worse than useless. Short reads cannot resolve long repeats or structural rearrangements at any coverage, which is why depth-based CNV segmentation has coarse breakpoint resolution ³. Base quality interacts with depth: a poor-quality 30X run genotypes worse than a clean 20X one ², so read both together. And no depth number tells you what a variant means clinically. Interpreting a specific finding for health decisions requires a clinician and, usually, orthogonal confirmation of the call.

Woolf Software builds longitudinal molecular profiles of individuals: whole-genome sequencing, RNA sequencing, proteomics, blood biomarkers, and continuous glucose data, integrated into one model of you. Build your profile.