Calling Mitochondrial Heteroplasmy: Why a Low-Level Variant Is as Likely to Be Nuclear DNA as Real

Zetobit Bioinformatics Insight Series

Mitochondrial Genomics

Calling Mitochondrial Heteroplasmy: Why a Low-Level Variant Is as Likely to Be Nuclear DNA as Real

The mitochondrial genome violates almost every assumption a standard variant caller is built on. The most consequential violation is invisible: fragments of mtDNA scattered across the nuclear genome leak reads back onto the mitochondrial reference, manufacturing low-heteroplasmy variants that were never in a mitochondrion at all.

By Kanna Nandakumar, PhD · Zetobit

Point a germline pipeline at chrM and it will run without complaint and hand you a VCF. That is the problem. Nothing in the output announces that the assumptions underneath the caller — two alleles per site, a linear reference, one genome per cell — are all false for the mitochondrion. The variants come out formatted correctly and sorted by position, and the low-level ones among them are a mix of real biology and artifact in a ratio you cannot read off the file.

Getting mitochondrial variant calling right is less about a better caller than about respecting how thoroughly this genome differs from the nuclear one. Three of its properties break the standard model, and one of them — the nuclear copies of the mitochondrial genome — is subtle enough that it routinely reaches clinical and population callsets disguised as genuine heteroplasmy.

Heteroplasmy is a continuous quantity, not a genotype

A nuclear variant is called as one of three states: homozygous reference, heterozygous, homozygous alternate. That trichotomy is baked into germline callers. But a cell carries hundreds to thousands of mtDNA copies, and a variant can be present in any fraction of them. Most variants are homoplasmic — present in essentially all molecules — but some are heteroplasmic, present in only a fraction, and that fraction is the clinically meaningful quantity.¹ Mitochondrial disease often shows a threshold effect, where a pathogenic variant must exceed some heteroplasmy level before disease manifests, so the number after the decimal point is not a nuisance parameter — it is the result.

This is why the field borrows a somatic caller. The gnomAD mtDNA pipeline uses a mitochondria mode of GATK MuTect2 — a caller originally designed to find sub-clonal variants in tumors — precisely because it estimates a continuous allele fraction rather than forcing a genotype.¹ A germline caller asked the wrong question of this genome will answer it anyway, quantizing a continuous signal into het-or-hom and discarding the very measurement you needed.

The reference is a circle pretending to be a line

The mitochondrial genome is circular; the reference is a linear string. The seam between the two — the non-coding control region that spans the artificial break at coordinates chrM:16024–16569 and chrM:1–576 — is exactly where reads struggle to map, because a read straddling the break has no contiguous home in the linear representation.¹ The gnomAD solution is deliberately physical: extract the chrM reads, realign them to a reference shifted by 8,000 bases, call the control-region variants on that shifted alignment, then convert coordinates back.¹ The break moves to the middle of nowhere, the control region becomes contiguous, and the variants that were unmappable at the seam become callable.

This is the same lesson coordinate work keeps teaching: the reference is a specific representation with its own artifacts, not a neutral coordinate space. A pipeline that treats chrM as just another short contig inherits the seam as a silent blind spot over the most variable, most clinically watched stretch of the genome.

NUMTs: the confound that scales with the wrong thing

The deepest problem is evolutionary. Over millions of years, pieces of the mitochondrial genome have integrated into the nuclear genome, leaving behind nuclear sequences of mitochondrial origin — NUMTs. Many are already in the reference assembly; others are polymorphic, present in some individuals and not others. Because a NUMT is a near-copy of mtDNA, reads arising from it can mis-align onto the mitochondrial reference and generate false-positive calls — and they do so at low heteroplasmy, right in the band where real low-level variants live.¹ The bias is not even symmetric: because mutated fragments preferentially align to their NUMT locus, the confound has been shown to skew heteroplasmy estimates toward the reference allele, so a real variant can be pulled below a threshold and lost while an artifact is invented elsewhere.³

Here is the counterintuitive part, and the reason this confound is so treacherous. Whether a NUMT-derived read looks like a plausible low-heteroplasmy variant depends on the ratio of nuclear to mitochondrial molecules in the sample — that is, on mtDNA copy number. A sample rich in mitochondria drowns the handful of nuclear-origin reads; a sample poor in mitochondria lets them rise to an apparent heteroplasmy that clears a naive threshold. In the gnomAD analysis, mtDNA copy number varied across cohorts and drove the fraction of NUMT-derived false positives, which could account for the majority of putative heteroplasmies.² The majority. The low-level variants in a careless callset are not mostly real with some noise — in the wrong sample they are mostly noise.

The same nuclear reads, two verdicts. A fixed heteroplasmy threshold does not protect against NUMTs, because the apparent fraction of a NUMT-derived variant depends on how many true mtDNA molecules are diluting it. The identical contaminating reads fall below the line in a mitochondria-rich sample and above it in a mitochondria-poor one.

Why a flat threshold is the wrong instinct

Faced with low-level noise, the reflex is to raise a heteroplasmy cutoff and move on. gnomAD did exactly this — and was explicit that it is a blunt instrument. For the v3.1 release the team filtered out variants below 10% heteroplasmy to avoid false positives from contamination, sequencing errors, and NUMT misalignment, while noting that later releases would try to recover sub-10% variants that pass additional QC.¹ A flat 10% floor is defensible for a population database whose job is aggregate frequencies. It is a poor fit for a clinical case where the pathogenic variant may genuinely sit at 5% — because the cutoff that suppresses the NUMT artifact also discards the real low-heteroplasmy call you were looking for.

The more durable defenses attack the cause rather than the symptom. The strongest is sample-level: because the confound scales with copy number, you exclude the samples where it dominates. gnomAD removed samples with estimated mtDNA copy number below 50 (over 6,500 samples) as prone to NUMT-derived false positives, and also removed samples above 500 as likely cell lines carrying culture-related somatic heteroplasmies.¹ Copy number, computed as roughly twice the mean mitochondrial coverage over the median autosomal coverage, is not a QC afterthought here — it is a primary filter, because it is the axis the artifact rides in on.

A fixed heteroplasmy threshold treats a variable problem as if it were constant. The apparent level of a NUMT artifact is set by the sample’s mitochondrial copy number, so the honest control is to condition on copy number — not to pick one percentage and apply it to every sample regardless of how much true mtDNA was there to dilute the noise.

Beyond copy number, robust mtDNA calling layers several more checks: excluding samples with contamination above a low threshold (gnomAD used a 2% contamination cutoff, combining nuclear freemix, haplogroup-based Haplocheck, and an internal NUMT-aware estimate); hard-filtering a short list of artifact-prone sites — positions 301, 302, 310, 316, 3107, and 16182 — where homopolymer tracts and a reference N at chrM:3107 manufacture apparent variants in everyone;¹ and flagging variants that are common at low heteroplasmy across the cohort, since those are enriched for shared error modes rather than shared biology. None of these is exotic. They are simply the price of admitting that chrM is a different genome.

What this means in practice

If you inherit a callset with mitochondrial variants in it, a few questions separate a trustworthy result from a plausible-looking one:

• Was a mitochondria-aware caller used? A germline caller on chrM quantizes heteroplasmy into het/hom and throws away the fraction. A somatic-style caller (MuTect2 mitochondria mode, or a dedicated tool like mity or MToolBox) estimates the level.²
• Is mtDNA copy number reported per sample? Without it you cannot judge how exposed a low-heteroplasmy call is to NUMT contamination — and copy number below roughly 50 is a red flag for exactly that.¹
• Was the control region handled specially? If nothing shifted the circular reference, variants near the chrM:16024–576 seam are suspect by omission.¹
• Are the known artifact-prone sites filtered? Calls at chrM:301/302/310/316/3107/16182 are technical until proven otherwise.¹
• Is the source tissue known? Cell lines accumulate their own heteroplasmies; a variant real in a lymphoblastoid line may not exist in the donor.¹

The mitochondrial genome is small enough to feel trivial — sixteen and a half kilobases, a single circular molecule, no introns. That smallness is a trap. Every property that makes it compact also makes it behave unlike the nuclear genome your tools were built for, and the low-heteroplasmy variants that matter most clinically are exactly the ones a naive pipeline is least equipped to tell apart from nuclear DNA wearing a mitochondrial disguise.

References

1. Laricchia K, Calvo SE, et al. gnomAD v3.1 Mitochondrial DNA Variants. gnomAD browser, November 2020. gnomad.broadinstitute.org/news/2020-11-gnomad-v3-1-mitochondrial-dna-variants
2. Laricchia KM, Lake NJ, Watts NA, et al. Mitochondrial DNA variation across 56,434 individuals in gnomAD. Genome Research. 2022;32(3):569–582. pmc.ncbi.nlm.nih.gov/articles/PMC8896463
3. Wei W, Chinnery PF, et al. NUMT confounding biases mitochondrial heteroplasmy calls in favor of the reference allele. Frontiers in Cell and Developmental Biology. 2019;7:201. frontiersin.org/articles/10.3389/fcell.2019.00201

© Zetobit LLC · Bioinformatics Insight Series zetobit.com