Sample Identity and Cross-Contamination: Why a Technically Perfect Pipeline Can Answer for the Wrong Person

Zetobit.

Bioinformatics Insight Series

Sample Identity and Cross-Contamination: Why a Technically Perfect Pipeline Can Answer for the Wrong Person

Coverage, duplication rate, mapping rate, transition-transversion ratio — every standard QC metric asks whether the data is good. None of them asks whose data it is. That question has its own failure modes, and they are invisible to everything else in the pipeline.

By Kanna Nandakumar, PhD · Zetobit

Most pipeline failures announce themselves. Coverage comes in low, duplication spikes, the mapping rate collapses, a caller throws an error. You investigate, you fix it, you rerun.

Identity failures do none of that. A swapped sample produces a beautifully clean BAM: 40× mean coverage, 99% mapping rate, unremarkable insert size distribution, a variant call set with entirely normal Ti/Tv. Every metric passes. The pipeline has answered a well-posed question with high confidence. It has just answered it about a different person.

This is the one class of error where the quality of your analysis provides no protection at all — where, if anything, a more polished pipeline makes the wrong answer more persuasive.

The error rate is not negligible

Sample identity errors are often treated as the kind of thing that happens to other laboratories. The published estimates suggest otherwise: across several clinical and research studies, the sample identity error rate has been estimated to range from roughly 0.2% to 6% in practice, spanning both complete swaps between samples and partial contamination of one sample by another.

A large clinical study puts a floor under this. In an analysis of 25,015 clinical samples genotyped with a 96-SNP tracking panel across a clinical sequencing network, precise SNP tracking found no sample swaps within the testing laboratories — but comparing genotype-predicted sex against the sex on the test requisition surfaced 110 inconsistencies, about 0.44% of samples. The causes were mostly upstream of the lab: roughly half were clerical errors, and about 40% were undetermined mix-ups that had occurred before the samples reached the genome centers. A smaller share reflected biology rather than error — transgender participants and stem cell or bone marrow transplant recipients, where a genotype-phenotype mismatch is expected.

Two things follow. First, a well-run laboratory can genuinely achieve near-zero internal swap rates — this is a solvable problem. Second, the laboratory is not where most of the errors happen. Identity checks that only validate the analytic phase are checking the part of the process that was already working.

Figure 1. The two identity failure modes fail in opposite directions but share one property: neither degrades the metrics a pipeline normally watches. A complete swap yields technically excellent data about the wrong individual. Partial contamination yields technically excellent data that is mostly right, with a minority of foreign reads that mimic low-frequency variants.

Index hopping: contamination the instrument creates

The most counterintuitive source of cross-contamination is not a pipetting mistake. It is the sequencer itself.

Illumina instruments using patterned flow cells with exclusion amplification (ExAmp) chemistry — HiSeqX, HiSeq 3000/4000, and NovaSeq — exhibit a phenomenon in which library fragments acquire the wrong sample index during clustering. The mechanism is chemical: residual free primer or adapter carried into a pooled library can hybridize and extend during the ExAmp step, producing a fragment tagged with another sample's index. Reads are then demultiplexed into the wrong sample's file.

The Broad Institute characterized this across two years of production data. They observed index swapping at a rate of 0.2% to 6% in all sequencing runs on those instruments, with the rate varying substantially by library preparation method. In their workflows, PCR-free whole genomes swapped at roughly 3% on average while PCR-plus genomes averaged around 0.25% — a difference they attribute to PCR-free libraries being diluted very little before sequencing, leaving a higher ratio of free adapter to library fragments.

Several details of their investigation are worth carrying forward. The problem worsened as multiplexing increased, rising when they moved from 8-plex to 24-plex pools. It is a property of the chemistry, not of a particular lab's technique: resequencing the same library pools on a MiSeq dropped contamination roughly ten-fold, which is what localized the event to the ExAmp flow cell process. And it is not confined to the sequencer — they observed elevated swap rates in exome libraries even on random-clustering instruments, because index swapping also occurs during the multiplex PCR of the pooled capture step itself.

Index-swapped reads are not errors. They are high-quality reads assigned to the wrong sample — which is precisely why quality filters do not remove them.

That distinction has teeth. The Broad team evaluating an RNA fusion caller found reads supporting the BCR–ABL1 fusion in the data files for all three cell lines they tested, when only K562 should carry that translocation. The average swap rate in their RNA-seq data was just 0.32%. A sub-1% contamination rate was enough to manufacture a clinically meaningful fusion call in two samples that did not have it.

The remediation is structural rather than computational. With single or combinatorial dual indexing, a swapped index still forms a valid-looking combination and demultiplexes silently. With non-redundant (unique) dual indexing, a swap produces an index pair that should not exist, and those reads can be filtered. The Broad reported that after implementing unique dual indexes across tens of thousands of exomes and genomes, average contamination rates fell close to zero. They also note a residual: rare "double swaps," where both indexes swap together, occurred at about 0.01% and cannot be filtered this way, since the resulting combination looks legitimate.

Why contamination is worse than it sounds at low percentages

A 1% contamination rate sounds tolerable, and for many germline applications it is. The Broad's own framing is that this level may not affect the ability to trust variant calling for many germline DNA applications — but it can produce spurious results when looking for rare transcripts or fusion events in RNA-seq, and in low allele fraction somatic analysis.

The reason is that contamination and the signal of interest live in the same numerical range. A somatic variant at 3% variant allele fraction, a subclonal resistance mutation, a ctDNA variant in a liquid biopsy, a rare fusion transcript — these are exactly the findings that a few percent of foreign reads can imitate. The contaminating reads carry a real individual's real genotype, so they are internally consistent, correctly paired, and well-mapped. They look like biology because they are biology, sourced from the wrong person.

This also means the tolerance for contamination is not a fixed number. It is a function of the lowest allele fraction the assay claims to detect. An assay validated to call variants at 2% VAF cannot coexist with a 3% contamination rate, regardless of how good every other metric looks.

Table 1. Identity and contamination failure modes by origin. The controls differ because the mechanisms differ — no single check covers the whole surface.

Failure	Where it originates	What detects it
Complete swap	Collection, accessioning, or labeling — often before the lab	Genotype concordance against an independent reference; sex concordance
Tumor–normal mispairing	Pairing the wrong normal to a tumor in a multi-sample case	Relatedness on germline sites tolerant of LOH
Cross-sample contamination	Pipetting, aerosol, shared plasticware, carryover	Contamination estimation from allele fractions at common SNPs
Index hopping	ExAmp clustering chemistry; also pooled capture PCR	Unique dual indexes with filtering of unexpected index pairs
Multi-assay mismatch	DNA and RNA from the same case diverging in the pipeline	Cross-data-type identity checks on shared SNP sites

What identity QC actually requires

The tooling here is mature and inexpensive relative to what it protects. The distinguishing feature of the modern tools is that they compare genotypes at a preselected panel of common SNPs — typically on the order of 20,000 sites — rather than requiring a jointly called variant set.

Genotype fingerprinting against an independent source

The strongest control is a genotype obtained by a different method than the assay being validated — an array, a SNP panel, or a prior sequencing run — compared against the sequencing data. Large programs build this in as standard practice: TOPMed's protocol describes molecular fingerprinting as part of initial sample QC, generating a unique genetic ID used to identify sample handling errors, with a fingerprint LOD score reviewed alongside coverage and contamination for each sample.

Contamination estimation as a routine metric

Tools such as VerifyBamID estimate the proportion of foreign DNA in a human sequencing sample from allele fractions at known SNPs. TOPMed applied a threshold below 3% for this metric. The important discipline is treating contamination as a first-class QC output reviewed for every sample, not a diagnostic reached for after something already looks wrong.

Relatedness tools that survive tumor biology

Cancer cases are the hard case, because reports often contain only somatic variants and loss of heterozygosity distorts genotypes. Somalier was built for this — it works directly on alignments without requiring jointly called germline variants, extracting a compact sketch of informative variation per sample so that hundreds of sketches can be compared in under a second. HYSYS takes a complementary approach, testing concordance of homozygous SNPs on the reasoning that homozygous germline population variants rarely change in the disease and are not affected by loss of heterozygosity.

Cross-data-type checks

When a case spans DNA and RNA, or tumor and normal, or multiple biopsies, each pairing is an opportunity for mismatch. NGSCheckMate was designed to verify identity across formats — FASTQ, BAM, or VCF — and across data types, including matching RNA-seq to exome data from the same individual, at sequencing depths as low as 0.5×. Peddy addresses the same problem from the pedigree side.

Sex concordance as a cheap first filter

Genotype-predicted sex against the requisition is nearly free and, as the 25,015-sample study showed, catches real errors that internal tracking misses. The caveat matters clinically: a mismatch is not automatically an error. Transplant recipients and transgender participants both produce legitimate discordance, and a workflow that treats every mismatch as a swap will generate its own problems.

The short version

Quality QC and identity QC answer different questions, and passing the first tells you nothing about the second. Identity checks belong in the pipeline as standing metrics with defined thresholds — reviewed for every sample, on every run, with the tolerance for contamination set by the lowest allele fraction the assay claims to report.

A question worth asking explicitly

What makes this failure mode distinctive is that it inverts the usual relationship between rigor and reliability. Elsewhere in bioinformatics, a more careful pipeline produces a more trustworthy result. Here, a more careful pipeline produces a more confident wrong answer, because every layer of polish adds credibility to a conclusion about the wrong individual.

The remedy is not more rigor in the same direction. It is asking a question that the rest of the pipeline structurally cannot ask: not "is this data good," but "does this data belong to the person whose name is on the report." Those are separate checks, and only one of them is usually running.

Zetobit builds and validates CAP/CLIA-compliant NGS pipelines, including sample identity and contamination controls designed around the detection limits an assay actually claims. If you are standing up identity QC or reconciling a suspected swap across a cohort, we're happy to talk.

References

1. Costello M, Fleharty M, Abreu J, et al. Characterization and remediation of sample index swaps by non-redundant dual indexing on massively parallel sequencing platforms. BMC Genomics 19:332 (2018). doi:10.1186/s12864-018-4703-0.
2. Venner E, Muzny D, Smith JD, et al. Genetic sex validation for sample tracking in next-generation sequencing clinical testing. BMC Research Notes (2024). doi:10.1186/s13104-024-06723-w.
3. Jun G, Flickinger M, Hetrick KN, et al. Detecting and estimating contamination of human DNA samples in sequencing and array-based genotype data. American Journal of Human Genetics 91(5):839–848 (2012). doi:10.1016/j.ajhg.2012.09.004.
4. Pedersen BS, Bhetariya PJ, Brown J, et al. Somalier: rapid relatedness estimation for cancer and germline studies using efficient genome sketches. Genome Medicine 12:62 (2020). doi:10.1186/s13073-020-00761-2.
5. Lee S, Lee S, Ouellette S, et al. NGSCheckMate: software for validating sample identity in next-generation sequencing studies within and across data types. Nucleic Acids Research 45(11):e103 (2017). doi:10.1093/nar/gkx193.
6. Sherry ST, et al. HYSYS: have you swapped your samples? Bioinformatics 33(8):1249–1251 (2017). doi:10.1093/bioinformatics/btw817.
7. TOPMed Whole Genome Sequencing Project. Sequencing methods and quality control documentation. dbGaP study documentation (2016).
8. Chen X, et al. PISAD: reference-free intraspecies sample anomalies detection tool based on k-mer counting. GigaScience (2025). doi:10.1093/gigascience/giaf061.