Mosaicism vs. Sequencing Artifact: Distinguishing True Low-VAF Variants from Noise

Zetobit · Bioinformatics Insight Series

Mosaicism vs. Sequencing Artifact: Distinguishing True Low-VAF Variants from Noise

At a 3% variant allele fraction, a real post-zygotic mutation and a library-prep chemistry error look almost identical in the pileup. Almost. The difference lives in signals most default pipelines discard before they ever reach interpretation.

By Kanna Nandakumar, PhD · Zetobit, LLC

A germline heterozygous variant sits near 50% VAF and a homozygous one near 100%. Those are comfortable numbers — well clear of the noise floor, easy to call with confidence. Mosaic variants live somewhere else entirely. A post-zygotic mutation that arose after the first few cell divisions is present in only a fraction of cells, so its allele fraction can be 5%, 2%, or lower, depending on when in development it happened and which tissue you sampled.¹ That is exactly the allele-fraction band where sequencing and sample-prep errors also live — and where a caller tuned for clean germline genotypes is least equipped to tell signal from noise.

The stakes are real. Mosaic variants underlie a growing list of conditions — focal epilepsies, overgrowth syndromes, some neurodevelopmental disorders — and a mosaic finding in a parent has very different recurrence-risk implications than a de novo call. Getting the call right means separating a handful of true alternate-allele reads from a background of look-alike artifacts. This piece lays out what actually distinguishes the two.

A germline het sits at a comfortable 50%. A mosaic variant might sit at 2% — in the same allele-fraction band where library-prep chemistry errors live. That overlap is the entire problem.

Why low-VAF calling is hard in the first place

Two forces converge in the sub-5% band. First, the true signal is genuinely faint: at 2% VAF and 200× coverage, a real mosaic variant is supported by roughly four reads. Random sampling alone makes that number wobble, and any locus-specific mapping trouble can erase it. Second, the background is not clean. A range of artifactual processes produce exactly this pattern — a small number of alternate-allele reads at a site that is truly homozygous reference.

The dominant artifact sources are well characterized. FFPE fixation deaminates cytosine to uracil, which reads as spurious C-to-T (and G-to-A) transitions; these can come to dominate the low-frequency call set in fixed tissue.² Oxidative damage — 8-oxoguanine, generated during acoustic shearing and handling — produces a characteristic G-to-T / C-to-A signature.³ On top of these, PCR polymerase errors, mapping artifacts near repeats, and sequencing miscalls all contribute their own low-level noise. None of these is biology. All of them can produce a clean-looking variant at 2–4% VAF.

The signals that separate them

The good news: true mosaic variants and artifacts differ in ways that are measurable — if you keep the right information. Four signals do most of the work.

Strand balance. A real variant is present on the original double-stranded DNA fragment, so reads covering it from the forward and reverse strands should both carry the alternate allele in proportion to their coverage. Many damage artifacts are single-stranded lesions and therefore appear preferentially on one strand — the basis of orientation bias. FFPE deamination and 8-oxoG both show this asymmetry, with artifactual alternate alleles skewed toward read 1 versus read 2, and tools exploit exactly that to filter them.^2,4 A candidate supported almost entirely by one strand is a red flag; balanced support is reassuring.

Substitution context. Because the major artifact classes have signatures — C>T for deamination, G>T/C>A for oxidation — a candidate whose change matches a known damage type, in a sample prone to that damage, deserves extra scrutiny. Picard's CollectSequencingArtifactMetrics and CollectOxoGMetrics quantify the global presence of these signatures so you know whether a run is oxidation- or deamination-prone before you trust any individual low-VAF call.³

Read position and mapping quality. Artifacts cluster: at read ends, in low-complexity or repetitive regions, and at sites with poor mapping quality. True variants are distributed as the genome dictates, not as the error process dictates. Position-in-read bias and mapping-quality filters remove a large share of false positives — which is also why these are precisely the features that modern mosaic callers learn from.

Consensus across molecules. The most powerful discriminator is molecular. Unique molecular identifiers (UMIs) tag each original fragment, letting you collapse all reads from one molecule into a consensus and discard errors that appear in only some copies. Duplex sequencing goes further: it requires a variant to appear on both strands of the original duplex, which eliminates single-stranded damage artifacts like 8-oxoG almost entirely.⁵ If your assay is built for genuinely low VAFs, this molecular error suppression is the foundation everything else rests on.

Four discriminating signals. A true mosaic variant and an artifact can share a VAF but rarely share all four properties. Robust mosaic calling weighs them jointly rather than gating on any one.

Where the tools stand

Because no single heuristic is sufficient, purpose-built mosaic callers integrate these signals — increasingly with machine learning. MosaicHunter and MosaicForecast pioneered non-cancer mosaic SNV detection, and DeepMosaic applies a convolutional network to an image-like representation of the pileup, achieving control-independent detection with strong reported precision on non-cancer whole-genome data.⁶ Each recovers true variants the others miss, which is why multi-caller pipelines have been common practice.⁶

The honest caveat is that low-VAF performance remains an active, unsolved problem. Recent benchmarking work makes the difficulty concrete: several established tools carry high false-positive counts in the low-VAF regime, and a substantial fraction of their false positives fall above even a 5% VAF threshold — the band that should be easiest.⁷ Newer approaches keep arriving to close the gap, including scalable bulk-sequencing methods and long-read callers that exploit haplotype phasing to separate genuine mosaics from artifacts.^7,8 The field is improving quickly, but "run one caller and trust the VCF" is not yet a safe posture at 2% VAF.

A practical protocol

The reliable path is to design against artifacts before calling and to weigh evidence after. Upstream, match the assay to the ambition: if you need VAFs below a few percent, use UMIs and consider duplex sequencing, and repair or avoid damage-prone input (FFPE repair enzymes, antioxidant handling) rather than filtering the damage out later. Profile every run with artifact metrics so you know whether it is oxidation- or deamination-prone. Downstream, use a caller built for mosaicism rather than a germline genotyper, and evaluate candidates on the full weight of evidence — strand balance, substitution context, read-position and mapping quality, and molecular consensus — not on VAF alone. Finally, treat orthogonal validation as the arbiter for anything clinically actionable: an independent method on an independent aliquot is what converts a promising low-VAF call into a reportable one. A real mosaic variant leaves a consistent story across all these axes. An artifact almost never does. ▪

---

References

1. Yang X, et al. Control-independent mosaic single nucleotide variant detection with DeepMosaic. Nature Biotechnology (2023) — mosaic variants as post-zygotic mutations with characteristically low, tissue-dependent allele fractions; reported sensitivity 0.78, specificity 0.83, PPV 0.96 on non-cancer WGS. nature.com/articles/s41587-022-01559-w
2. Khan MS, Shih DJH. Filtering sequencing artifacts from FFPE tissues: an orientation-bias-based statistical tool (MOBSNVF). Briefings in Bioinformatics (2025) — FFPE cytosine deamination produces C:G>T:A artifacts that can dominate low-frequency calls; orientation bias (read 1 vs read 2) distinguishes them. ncbi.nlm.nih.gov/pmc/articles/PMC12699713
3. bcbio damage-filter documentation, summarizing DKFZ bias filtering and Picard artifact metrics — oxidative (8-oxoG) damage yields a G>T (read 1) / C>A (read 2) signature; CollectOxoGMetrics and CollectSequencingArtifactMetrics quantify global damage. github.com/bcbio — damage filters
4. Comprehensive simulation of tumour sequencing data — Mutect2's optional orientation-bias filter removed ~99% of FFPE/8-oxoG artifacts that passed standard filters, though a residual fraction still slipped through as PASS. NAR Cancer (2023). academic.oup.com/narcancer/article/5/3/zcad051
5. High-efficiency error suppression for low-frequency variants — UMI single-strand consensus removes per-read errors; strand-aware duplex consensus retains only variants found on both strands of a fragment, eliminating oxidative-damage artifacts. Nucleic Acids Research (2019). academic.oup.com/nar/article/47/15/e87
6. Yang X, et al. DeepMosaic (as above) — CNN-based, control-independent mosaic SNV classifier; documents variants uniquely recovered relative to MosaicForecast, MosaicHunter, MuTect2, and Strelka2, motivating multi-caller pipelines. biorxiv.org/content/10.1101/2020.11.14.382473v2
7. Illumina. Scalable Mosaic Variant Detection at Low Allele Fractions with DRAGEN — benchmarking on GIAB HG002 mosaic sets showing high false-positive counts for established callers in the low-VAF regime, with large fractions of their false positives above 5% VAF. developer.illumina.com — DRAGEN mosaic
8. Clair-Mosaic — deep-learning long-read mosaic caller that leverages the haplotype relationship of phased long reads to distinguish genuine mosaic variants from sequencing artifacts. bioRxiv (2025). biorxiv.org/content/10.1101/2025.10.31.685831v1

ZETOBIT.COM · BIOINFORMATICS INSIGHT SERIES