Finding. Shotgun metagenomic pipelines throw the host away. In low-biomass swabs that can be more than 90% of the run. Tomar and Khairnar kept those discarded human reads from 47 upper-respiratory swabs collected from healthy adults in Nagpur, mapped them to the revised Cambridge Reference Sequence, and recovered a complete mitochondrial genome from every sample. Mean depth was 433-fold (range 9-fold to 3,029-fold). Haplogroups landed in 46 of 47 people (macrohaplogroups M 22, R 23, N 1). The hard result is the allele-fraction floor. After nuclear mitochondrial segment cleanup and a microbial filter, heteroplasmy above 5% looks like biology. Below 5%, a T>A error process eats the call set, and a substitution-specific beta-binomial plus a leave-one-out panel of normals does not save you.
Why this paper matters
Mitochondrial DNA heteroplasmy is how pathogenic variants hide: a mutant genome at 2% today can become 60% in a daughter tissue tomorrow. The clinical instinct is to push allele fractions as low as the sequencer allows. This preprint is a stop sign for one increasingly common data exhaust. Respiratory metagenomes are already being generated for infection, microbiome, and surveillance work. If leftover host reads are a free mitochondrial genome, they will be mined. If they are a free error generator, they will contaminate heteroplasmy databases.
The authors do not sell a miracle caller. They sell a limit. Competitive realignment to GRCh38 moved 3.22% of rCRS-recruited reads onto nuclear loci. That is the NUMT problem in a single number: a few percent of "mitochondrial" reads were nuclear lookalikes. Microbial cross-mapping via Kraken2 removed only 0.045%, so bacteria are not the main liar here. The liar is chemistry plus NUMTs.
What they actually measured
Error rates across substitution classes varied 12.8-fold, and T>A was the worst. The raw call set (134 sites) had a transition/transversion ratio of 0.43. In gnomAD v3.1 that ratio is 3.51. T>A alone was 42.5% of calls and never rose above 4.72% allele fraction, while genuine-looking transitions reached 90.81%. Raise the threshold to 5% and the T>A pile disappears, Ti/Tv climbs to 2.20 and plateaus, and 16 calls remain. That is the paper's operating rule: leftover host reads support complete mitochondrial recovery, haplogroup typing, and heteroplasmy detection for allele fractions above 5%.
How to read the score
High 80s. This is core mitochondrial genetics with a number you can put in a pipeline. Confidence is high for the technical claim in this sample type (healthy-adult upper-respiratory shotgun data). It is not a patient series and not a chemistry paper that names the T>A adduct.
Caveats
Healthy people only. One swab type. Depth includes a 9-fold genome, where a 5% variant is a handful of reads. The residual T>A process is characterized, not explained. Do not cite this brief as proof that 1% pathogenic heteroplasmy is undetectable in blood or muscle. Do cite it as proof that background-fitted models are not a license to believe every low-allele T>A.
What to do with it
If you already have host-depleted metagenomes, this is permission to rebuild mtDNA and haplogroups, plus a rule for what not to publish as heteroplasmy. Pull the workflow (GitHub tomar-neeri/mtdna-heteroplasmy-host-reads, Zenodo 10.5281/zenodo.23034211) and the BioProject PRJNA1478410 reads. If you build clinical mtDNA callers, add a substitution-class dashboard and a hard look at T>A below 5% before you claim mosaic disease.