False Positives in Metagenomic Classification: Why a Handful of Reads Can Invent a Pathogen That Was Never There

Zetobit Bioinformatics Insight Series

Clinical Metagenomics

False Positives in Metagenomic Classification: Why a Handful of Reads Can Invent a Pathogen That Was Never There

In metagenomic pathogen detection the organism you are hunting may be fewer than a hundred reads out of tens of millions — and contamination, a conserved sequence, or a dirty database entry can produce exactly that same handful. Read count alone cannot tell the real signal from the manufactured one.

By Kanna Nandakumar, PhD · Zetobit

Metagenomic classification takes millions of sequencing reads from a sample and assigns each one a taxonomic identity by matching short words — k-mers — against a database of reference genomes. Because there are far too many reads to align, classifiers such as Kraken do exact k-mer matching, which is fast enough to be practical and specific enough to be useful. For surveying the bacterial makeup of a gut or soil sample, where the organisms of interest are abundant, this works well.

Clinical metagenomics inverts the situation and, with it, the entire risk profile. Here the goal is to sequence an infected human sample directly and find the pathogen without culturing it — a powerful idea when the causative organism is unknown or unculturable. But the pathogen is now a needle in a haystack of human DNA. Typically 95–99% of reads match the host, and sometimes fewer than 100 reads out of many millions match the target organism.¹ When the true signal is that faint, anything that produces a comparable number of misassigned reads can imitate it — and several things routinely do.

Why read count is the wrong evidence

The instinctive metric is read count: more reads assigned to a species means more confidence it is present. In the low-abundance regime this instinct is actively dangerous, because there are at least three ways to accumulate a small pile of reads for an organism that is not in the sample.

The first is contamination. Reagents, extraction kits, and the laboratory environment carry their own microbial DNA, and this is not a rare edge case — it is a well-documented, reproducible feature of low-biomass sequencing. Sequence contamination introduced from kits or the environment during extraction, handling, or sequencing can yield high numbers of spurious identifications, and small amounts of input material compound the problem.² In a clinical sample, common skin bacteria from the patient or lab personnel can easily generate a read count similar to the target and thereby mask or mimic the real signal.² When the pathogen is 50 reads, a contaminant at 50 reads is indistinguishable by count.

The second is the database itself. Reference assemblies are not clean: many contain segments of contaminating sequence from other organisms, mislabeled at the source. Spurious matches arise from low-complexity regions of genomes and from contamination in the database genomes themselves, so a read can be confidently assigned to a species whose reference merely happens to contain a stretch of someone else’s DNA.¹ The classifier is faithfully reporting a match; the match is an artifact of a dirty reference.

The third is shared and low-complexity sequence. Conserved regions — ribosomal RNA genes are the classic example — and low-complexity or convergently evolved stretches attract reads from many organisms. Reads can be misassigned to highly conserved or convergently evolved genomic regions, producing identifications driven by sequence similarity in non-specific regions rather than by the organism’s actual presence.³

Breadth of coverage separates real from fake

The key insight that rescues the situation is that these false-positive reads share a geometry: they land in the same few places. A contaminant stretch in a database genome, a conserved rRNA operon, a low-complexity repeat — each is a small, specific target, so the reads it attracts pile up there and nowhere else. Such false-positive reads typically match only small portions of a genome; if the only reads matching a species fall in one low-complexity region, that species was probably not present.¹ A microbe that is genuinely in the sample behaves differently: its reads distribute relatively uniformly across the whole genome rather than concentrating in one or a few locations.¹

So the discriminating question is not “how many reads?” but “how much of the genome do they cover?” — depth versus breadth. Fifty reads spread across fifty different regions of a genome is strong evidence of presence; fifty reads stacked on a single conserved locus is evidence of an artifact. Read count conflates these two situations; breadth of coverage tells them apart.

Same count, opposite verdicts. A read total says nothing about where the reads land. Clustering on a single conserved or low-complexity region — covering few of the genome’s unique k-mers — is the signature of an artifact; a real organism scatters reads across the whole genome and lights up many unique k-mers. This is exactly the distinction that unique-k-mer counting operationalizes.

Counting unique k-mers, cheaply

Turning breadth into a fast, routine metric is what tools like KrakenUniq added on top of Kraken. Rather than counting reads assigned to a taxon, KrakenUniq counts how many of each genome’s unique k-mers are covered by the reads, using the HyperLogLog cardinality estimator to do so at Kraken’s speed and with little extra memory.¹ A pile of reads on one repeat covers the same few k-mers over and over and yields a low unique-k-mer count; a genuinely present organism covers many distinct k-mers across its genome. The measure approximates breadth of coverage without the cost of full alignment.

The practical payoff is concrete. In the original work, for discovering pathogens in human patients, a unique-k-mer count threshold of 1000 eliminated many background identifications that a read-count threshold would have let through.¹ That single number — unique k-mers, not reads — is often the difference between a report cluttered with laboratory bystanders and one that surfaces the actual pathogen.

Depth tells you how many times you saw a sequence; breadth tells you how much of the organism you saw. In low-biomass clinical metagenomics the artifacts are deep and narrow — many reads, tiny footprint — while real organisms are broad. Filtering on breadth (unique k-mers) rather than depth (read count) is the single most effective way to suppress false positives.

The defenses that actually work

Unique-k-mer filtering is the sharpest tool, but robust pathogen detection layers several defenses, because the false positives come from several sources:

• Remove host reads first, and thoroughly. With 95–99% of reads from the human host, aggressive host depletion both speeds classification and cuts false positives; classifiers that index the human genome in their database had negligible misclassification, making host inclusion a first important step to reducing false positives.⁴
• Filter on breadth, not just depth. Require a minimum unique-k-mer count (breadth of coverage) in addition to a minimum read count (depth); the KrakenUniq threshold of ~1000 unique k-mers is a documented starting point for pathogen discovery.¹
• Run negative controls and track known contaminants. Because kit and environmental contamination is reproducible, a no-template control processed identically reveals the background; an organism present at similar levels in the control is suspect.²
• Mind the database. Contaminated or mislabeled reference assemblies create false positives that no sample-level control catches; a curated, host-inclusive database matters as much as the classifier.¹
• Distrust identifications built on conserved regions. A call resting on rRNA or other conserved, low-complexity sequence is weak evidence regardless of read count.³

The through-line is the same one that recurs whenever a measurement is pushed to its detection limit: at high abundance the naive metric is fine, and at low abundance it quietly inverts. A read count that would be perfectly adequate for profiling a gut microbiome becomes a generator of phantom pathogens when the true signal is a hundred reads in ten million. The organisms that matter most in a clinical sample are exactly the ones sitting closest to the noise — and telling them apart from the noise is a question of where the reads fall, not how many there are.

References

1. Breitwieser FP, Baker DN, Salzberg SL. KrakenUniq: confident and fast metagenomics classification using unique k-mer counts. Genome Biology. 2018;19:198. link.springer.com/article/10.1186/s13059-018-1568-0
2. Breitwieser FP, Baker DN, Salzberg SL. KrakenHLL: confident and fast metagenomics classification using unique k-mer counts (preprint; sample and reagent contamination). bioRxiv. 2018;262956. Citing Salter et al. (2014) and Thoendel et al. (2017). biorxiv.org/content/10.1101/262956v2.full
3. Refining filtering criteria of the Kraken family of tools for accurate taxonomic profiling of ancient metagenomic data. PMC. 2025. pmc.ncbi.nlm.nih.gov/articles/PMC13223122
4. Ye SH, Siddle KJ, Park DJ, Sabeti PC. Benchmarking metagenomics tools for taxonomic classification. Cell. 2019;178(4):779–794. sciencedirect.com/science/article/pii/S0092867419307755

© Zetobit LLC · Bioinformatics Insight Series zetobit.com