Coverage Uniformity vs. Mean Depth: Why the Average That Reassures You Hides the Dropout That Fails the Sample

Zetobit · Bioinformatics Insight Series

Coverage Uniformity vs. Mean Depth: Why the Average That Reassures You Hides the Dropout That Fails the Sample

"Sequenced to 100× mean coverage" sounds like an assurance that every base was read a hundred times. It isn't. Mean depth is an average over a distribution, and two samples with the same average can differ enormously in whether any given clinically important base was covered at all — because sequencing dropout is systematic, not random.

By Kanna Nandakumar, PhD Zetobit LLC

Depth of coverage is the headline number of nearly every sequencing report: 30× for a genome, 100× or more for a clinical exome or panel.¹ It is intuitive, easy to quote, and central to how labs specify and price their work. It is also, on its own, one of the more misleading summaries in genomics — because it is a mean, and a mean says nothing about how the reads are distributed across the target.

A sample sequenced to 100× mean might have every base sitting comfortably between 80× and 120×, or it might have half its bases at 190× and the other half at 10× — or a set of clinically important exons at zero. All three report the same mean. The average reassures; the distribution is where the variant calls live or die. And the reason this matters so much is that the low-coverage regions are not scattered at random. They are the same regions, sample after sample, driven by predictable properties of the sequence.

Dropout is systematic, and GC content is the main driver

The dominant cause of uneven coverage is GC bias: the dependence of read depth on the local GC content of a fragment. The effect is unimodal — both GC-rich and AT-rich regions are underrepresented relative to those in the roughly 40–60% GC "sweet spot," with coverage falling off toward both extremes.² Much of this originates in PCR during library prep, which amplifies moderate-GC fragments more efficiently than extreme ones, compounded by capture and sequencing chemistry.³ GC-rich regions such as CpG islands and promoters can also form stable secondary structures that resist amplification and sequencing, leaving gaps.⁴

The consequence is that dropout has a genomic address. Systematic analysis has shown that many transcription start sites and first exons in the human genome are poorly covered because they are GC-rich — enough so that researchers catalogued a list of roughly a thousand "bad promoters" that resist sequencing on standard platforms.⁵ Coverage evenness is further shaped by exon size, repeat content, and segmental duplications, and uniformity tends to decrease for longer exons.⁶ None of this is visible in a mean. All of it is reproducible, which means a mean-depth spec silently and consistently under-serves the same genes in every sample that uses it.

Both distributions average to 100×. The uniform one keeps every base above the callable threshold; the uneven one piles extra depth on easy regions while a subpopulation of hard bases falls below 20× into un-callable territory. The mean is identical; the clinical result is not.

The metrics that actually capture uniformity

Because the mean hides the distribution, uniformity needs its own metrics — and several exist precisely for this. The most direct for clinical work is the percent of target bases covered at or above a callable threshold (for example, ≥20× or ≥30×). For confident exome variant calling, a common expectation is that at least ~90% of target bases reach 20×, and this number, unlike the mean, penalizes dropout directly.⁷

The Fold-80 base penalty captures uniformity in a single number: the fold of additional sequencing that would be required to bring 80% of target bases up to the mean coverage. A perfectly uniform run scores 1.0; higher values mean more unevenness, with a common clinical threshold of ≤1.4 (ideally ≤1.2).⁸ A related evenness score expresses uniformity as a percentage, where 100% is perfectly even.⁹ Each condenses the shape of the distribution into something the mean cannot convey.

Even the uniformity metrics have blind spots. Fold-80, as classically computed, emphasizes low-coverage regions but omits zero-coverage regions from the calculation — so complete dropout, the most dangerous case, can be underweighted by the very metric meant to catch unevenness.¹⁰ This is why no single number suffices: percent-above-threshold catches zero-coverage bases that Fold-80 may miss, and Fold-80 characterizes the spread that a threshold count doesn't. They are complementary, not interchangeable.

Why this is a clinical safety issue, not a nicety

The stakes are concrete. In exome data with high average depth (over 75×), specific regions can still be captured at as little as 10×, low enough to compromise variant detection — and sensitivity for detecting variants drops substantially in low-coverage regions relative to well-covered ones.⁶ A variant that is genuinely present but sits in a dropout region simply won't be called, and the high mean depth on the report gives false confidence that the negative result is trustworthy.

This bias even distorts which genes get studied. An analysis of clinically important cardiac genes found individual gene coverage to be markedly non-uniform and significantly associated with GC content — with the concern that disease causation gets preferentially attributed to well-covered genes simply because the poorly-covered ones are systematically under-interrogated.¹¹ A "negative" exome may mean the causal gene was never adequately read, not that no causal variant exists. Mean depth cannot distinguish those two outcomes; per-region coverage can.

What a defensible coverage report requires

Treating coverage as a distribution rather than a single average means reporting and checking what the mean conceals:

• Report percent-above-threshold, not just the mean. State the fraction of target bases at ≥20× (and/or ≥30×) as the primary adequacy metric, since it penalizes dropout directly.⁷
• Include a uniformity metric. Add Fold-80 penalty or an evenness score so the shape of the distribution is visible, not just its center.⁸
• Track coverage per region, per gene. For clinical panels/exomes, evaluate coverage at the level of the genes and exons that matter, not as a single genome-wide summary.¹¹
• Report callable regions explicitly. Distinguish "no variant found" from "not adequately covered" by flagging under-covered targets, so a negative result carries its coverage caveat.⁶
• Watch GC and design for it. Monitor GC-bias metrics and choose library/capture chemistry that minimizes it, since the worst dropout is predictable from sequence content.³

As with the other blind spots in this series, the trustworthy report states what the summary number hides: not just mean depth, but the fraction of targets callable, the uniformity of the distribution, and which specific regions fell short. A sample described only by its mean looks well-sequenced, but the average is precisely the statistic that cannot tell you whether the base you care about was read at all.

The takeaway

Mean depth is the most quoted and least sufficient coverage metric in sequencing. It averages over a distribution whose low tail — not its center — determines whether variants can be called, and that low tail is systematic: GC-extreme regions, promoters and first exons, long exons, and repeats drop out in the same places every run. Uniformity metrics like percent-above-threshold and Fold-80 exist to expose what the mean conceals, though each has its own blind spot, so none should stand alone. The discipline is to stop treating a high average as an all-clear and to ask the question the mean can't answer: not how deep on average, but how evenly — and was the base that matters covered at all?

References

1. Whole-exome sequencing coverage requirements — typical clinical WES 100–200× mean; depth, breadth, Fold-80 penalty ≤1.4 (ideally ≤1.2) as QC. CD Genomics. cd-genomics.com
2. Benjamini & Speed — GC content bias is unimodal: both GC-rich and AT-rich fragments underrepresented; full-fragment GC drives fragment count; PCR the major cause. Nucleic Acids Res. 2012. academic.oup.com
3. GC/PCR bias — PCR preferentially amplifies ~40–60% GC fragments; extremes underrepresented; capture and sequencing chemistry compound it. ScienceInsights / Revvity. 2025. revvity.com
4. GC-rich regions (CpG islands, promoters) form stable secondary structures that hinder amplification/sequencing → underrepresentation and gaps. Revvity. 2025. revvity.com
5. Ross MG, et al. Characterizing and measuring bias in sequence data — GC-rich first exons/transcription start sites poorly covered; ~1,000 "bad promoters" resistant to sequencing. Genome Biol. 2013. genomebiology.com
6. Novel metrics reveal local/global non-uniformity in WES — even at >75× mean, some regions covered ~10×; longer exons less uniform; low-coverage regions have reduced variant sensitivity. Sci Rep. 2017. nature.com
7. Coverage breadth — ≥90% of target bases at ≥20× expected for confident WES variant calling; % target at threshold as adequacy metric. Diagnostech NGS tips. diagnostech.co.za
8. Fold-80 base penalty — fold of additional sequencing to bring 80% of target bases to mean; 1.0 = perfect uniformity; higher = less uniform. Roche/IDT. diagnostics.roche.com
9. Evenness score — uniformity as a percentage (100% = perfectly even); complements Fold-80 in exome QC. PMC (clinical exome evaluation). 2023. PMC10138641
10. IDT — Fold-80 emphasizes low-coverage regions but omits zero-coverage regions; % of target bases within 50–200% of mean as a complementary measure. IDT white paper. idtdna.com
11. High-throughput exome coverage of cardiac genes — individual gene coverage non-uniform, significantly associated with GC content; risk of bias toward well-covered genes. PMC. PMC4272796

Zetobit LLC · Bioinformatics Insight Series · Lexington, KY · zetobit.com