The Sample You Think You Sequenced

Zetobit · Bioinformatics Insight Series

The Sample You Think You Sequenced

A variant call is only as trustworthy as the assumption that the reads came from one person — the right person. Contamination and sample swaps quietly break that assumption, and by the time they surface in a report, the damage is already done.

By Kanna Nandakumar, PhD · Zetobit, LLC

Somatic and germline pipelines are built on an unstated premise: the reads in this BAM come from the individual named on the requisition, and from no one else. Most of the time that holds. When it doesn't, the failure is silent. There is no error message for a tube mix-up, no warning flag when a neighboring well cross-contaminates during library prep, and no alignment artifact that announces "these two percent of reads belong to someone else." The pipeline runs cleanly, produces a VCF, and hands you variants that look exactly like real biology.

That is what makes contamination detection a first-class quality-control step rather than an optional add-on. It comes in three flavors — cross-sample contamination between humans, sample swaps, and cross-species contamination — and each distorts a clinical call in a different way.

There is no alignment artifact that announces "two percent of these reads belong to someone else." The pipeline runs cleanly and hands you variants that look exactly like real biology.

Cross-sample contamination: when a little is a lot

Cross-sample (cross-individual) contamination is the presence of DNA from another person mixed into the intended sample. The clinical problem is asymmetric: even a small contaminant fraction can matter enormously. The contaminant's germline heterozygous sites show up as low-frequency alternate alleles in your data — and a somatic caller, whose entire job is to find real low-frequency variants, has no way to know these particular low-VAF signals are foreign. The result is false-positive somatic calls, reduced specificity, and a contamination floor that sets how low a true variant allele fraction you can trust.¹

The workhorse for measuring it from sequence alone is VerifyBamID2. It takes an aligned BAM and a set of known variant sites with population allele frequencies, and models the observed reads as a mixture of the intended sample and a contaminant, reporting an estimated contamination fraction (the FREEMIX value).² Its key advance over the original method is being ancestry-agnostic: earlier approaches that assumed pooled or mismatched population allele frequencies systematically underestimated contamination — in the original validation, by margins ranging from roughly a fifth to nearly three-quarters depending on the population — because the prior genotype probabilities were wrong for that individual.³ VerifyBamID2 jointly estimates the sample's genetic-ancestry coordinates and the contamination level, so the estimate holds up across diverse and admixed ancestries.³

The number it produces is directly actionable. Because contamination has to sit well below your minimum allele fraction of interest, a sample with ~1% contamination may force you to disregard variants below roughly 5% VAF to stay safe.¹ That is not an abstract threshold — it is the line between reporting a subclonal driver and suppressing it, or between calling noise and calling a mutation.

Tumor–normal pairs and the swap problem

Cancer sequencing adds two complications. First, tumors carry copy-number changes that distort allelic ratios, which confounds contamination estimators built for diploid germline data. Second, the matched-pair design introduces a specific, insidious failure mode: the tumor and normal don't actually come from the same person — a sample swap.

Conpair was built for exactly this pair-aware setting. It does two things at once: it estimates cross-individual contamination in a way that remains reliable even in the presence of copy-number changes, and it checks concordance between the tumor and its matched normal to confirm they share an identity.⁴ On in-silico contamination ladders it resolves contamination down to about 0.1%, and its authors showed that estimated contamination in real TCGA glioblastoma tumor–normal datasets tracked closely with tumor–normal concordance and with the number of germline variants being miscalled as somatic by standard callers.^4,5 That last point is the whole argument: contamination isn't a cosmetic QC blemish; it directly inflates the somatic call set.

Why swaps deserve their own vigilance: if the "normal" belongs to a different individual, every germline variant that person carries and the patient doesn't will be subtracted incorrectly — or, worse, the patient's own germline variants will be called as somatic because they're absent from the mismatched normal. A swap doesn't degrade the result gracefully; it can invert it. Concordance checks that compare genotypes at a panel of markers between the two BAMs are the guardrail, and they belong in the pipeline before any somatic subtraction runs.

Three failure modes, three guardrails. Each contamination type is detected by a different signal and threatens a clinical call differently. None is caught by ordinary alignment QC — each needs a dedicated check.

Cross-species contamination: the reads that aren't human

The third flavor is non-human DNA in the sample. The canonical case is the patient-derived xenograft (PDX), where a human tumor grows in a mouse host and the extracted material inevitably contains mouse stromal DNA. Because the human and mouse genomes are highly homologous in many regions, a meaningful fraction of mouse reads will map to the human reference⁶ — and once there, they masquerade as human sequence, skewing allele fractions and generating spurious variants.

The fix is deconvolution rather than a single scalar estimate. Reads are compared against both host and graft genomes and assigned by origin — either by aligning to each reference and comparing edit distances (as tools like XenofilteR do), or by alignment-free k-mer classification directly on the FASTQs.^6,7 Done well, this removes the large majority of mouse-origin reads while retaining human sequence, which restores accurate variant allele fractions and recovers the true somatic mutations.⁷ Benchmarking work across PDX profiling has made clear that the choice of tool and reference materially affects downstream calls, so this is a step to validate, not to accept on defaults.⁸ Beyond xenografts, the same read-classification logic underlies detection of microbial or reagent contamination that would otherwise inflate background noise.

Why it matters clinically — and what to do

Every one of these failure modes shares a property that makes it dangerous: it produces a plausible, well-formed result. A contaminated sample yields extra somatic calls that look like subclonal biology. A swapped normal yields a somatic set that is confidently, systematically wrong. A PDX with unremoved mouse reads yields allele fractions that don't correspond to the patient's tumor at all. In a research setting these cost time and reagents. In a clinical setting — where a variant might gate a targeted therapy, a germline finding might prompt cascade testing in a family, or a tumor–normal comparison might define what gets reported — they cost more than that.

The practical posture is straightforward. Run contamination and identity checks as gates, before variant interpretation, not as a post-hoc audit. Estimate cross-sample contamination on every sample and treat the fraction as a hard input to your minimum reportable VAF. For tumor–normal work, add a pair-concordance check so a swap can never pass silently. For any xenograft or mixed-species material, deconvolute reads by origin and validate the tool against known mixtures. And monitor these metrics as trends across batches — a rise in contamination localized to one plate or one operator is a stop-the-line signal, not a number to normalize away. The sample you think you sequenced and the sample you actually sequenced are usually the same. Contamination detection is how you earn the right to assume it. ▪

---

References

1. Illumina DRAGEN Documentation. QC Metrics and Coverage/Callability Reports — cross-sample contamination modeling; requirement that contamination sit well below the minimum allele frequency of interest for reliable somatic calling. support-docs.illumina.com/SW/dragen_v42
2. Griffan / VerifyBamID2. A robust tool for DNA contamination estimation from sequence reads using an ancestry-agnostic method — inputs (BAM + population-AF VCF), .selfSM output, and the FREEMIX contamination estimate. github.com/Griffan/VerifyBamID
3. Zhang F, et al. Ancestry-agnostic estimation of DNA sample contamination from sequence reads. Genome Research (2020). Demonstrates underestimation of contamination under mismatched/pooled allele frequencies and the joint ancestry–contamination model. pubmed.ncbi.nlm.nih.gov/31980570
4. Bergmann EA, Chen B-J, Arora K, Vacic V, Zody MC. Conpair: concordance and contamination estimator for matched tumor–normal pairs. Bioinformatics 32(20):3196–3198 (2016). Detects sample swaps and cross-individual contamination; reliable at contamination as low as 0.1% despite copy-number changes. ncbi.nlm.nih.gov/pmc/articles/PMC5048070
5. Conpair, continued — estimated contamination in TCGA glioblastoma WGS/WXS tumor–normal data correlated with tumor–normal concordance and with germline variants miscalled as somatic by widely used callers. bioinformatics.nygenome.org — Conpair
6. Callari M, et al. Computational approach to discriminate human and mouse sequences in patient-derived tumour xenografts — high human–mouse homology causes a fraction of mouse reads to map to the human reference; comparison of dual-genome assignment approaches. PMC5755132. pmc.ncbi.nlm.nih.gov/articles/PMC5755132
7. Kluin RJC, et al. XenofilteR: computational deconvolution of mouse and human reads in tumor xenograft sequence data — edit-distance classification removes >99.9% of mouse-origin reads while retaining human sequence, restoring accurate VAFs and recovering somatic mutations. XenofilteR, BMC Bioinformatics (2018)
8. Bhandari M, et al. Benchmarking mouse contamination removing protocols in patient-derived xenografts genomic profiling. npj Precision Oncology (2025) — tool and reference choice materially affect downstream calls after mouse-read removal. nature.com/articles/s41698-025-00902-z

ZETOBIT.COM · BIOINFORMATICS INSIGHT SERIES