Is every sample from the person on its label?
Tubes get swapped, labels get typed wrong, and sample sheets slip by a row. In a study where each person is sampled more than once, one mix-up quietly breaks every paired test. SwapSeq reads each person's microbial fingerprint and points to the samples that belong to someone else.
- Swapped tubes
- Wrong labels
- Sample-sheet shifts
Tested on 4,732 samples from five public studies. In one of them it found a real tube swap without being told where to look.
Check your study
Open your feature table and the sheet that says who each sample came from. Both stay on your computer: SwapSeq runs inside this page. It works best with raw counts at ASV or OTU level and two or more samples per person.
Feature table
QIIME 2 artifacts, BIOM 1 and 2, and text tables all work. The file is read here in your browser and is not sent anywhere.
Sample metadata
First column: sample ID. One column: who the sample came from. The QIIME 2 format works as it is.
Check
- Feature table not opened yet
- Sample metadata not opened yet
Household and plate columns (optional)
Results
How it works
Three checks, all run in your browser on your own table.
A fingerprint of rare taxa
Everyone carries a few uncommon ASVs that stay with them week after week. SwapSeq compares which ASVs each sample holds, at a matched sequencing depth, and their log-ratio profile. A sample closer to somebody else than to its own person stands out.
A test for swapped pairs
Two swapped tubes leave a signature: tube A fits person B and tube B fits person A. SwapSeq tests every pair and names the partner when swapping their labels back makes both fit again.
Calibrated on your own table
Before judging, SwapSeq swaps hundreds of labels in a copy of your data to learn what an error looks like there. It then gives strong flags with almost no false alarms, a list to check against lab records, and an estimate of how many swaps it would catch in your study.
Tested on public studies
Known errors were planted in real published tables at about a 1% rate, 30 to 60 times each, and every result was counted. Bray-Curtis is the same pipeline with the distance used by a recent shotgun mislabel detector. Read the full methods and validation.
A real swap, found blind. In the Minnesota diet study (Johnson et al. 2019) the authors found and corrected two swapped tubes. With their original labels restored, SwapSeq reported exactly that pair as a swap at probability 1.00.
Show the full results table
| Dataset | Samples | People | Swaps flagged | + check list | Bray-Curtis | Mislabels flagged | + check list | False flags / 1000 | False checks / 1000 |
|---|---|---|---|---|---|---|---|---|---|
| SMP gut | 638 | 75 | 92% | 97% | 23% | 77% | 88% | 0.1 | 0.4 |
| SMP tongue | 692 | 80 | 43% | 61% | 3% | 13% | 60% | 0.1 | 2.8 |
| SMP palm | 515 | 61 | 52% | 82% | 2% | 10% | 47% | 0.4 | 2.2 |
| SMP forehead | 675 | 80 | 89% | 98% | 38% | 62% | 68% | 0.3 | 1.3 |
| ECAM infants | 875 | 43 | 2% | 24% | 0% | 0% | 25% | 0.3 | 14.6 |
| SMP gut (2 per person) | 150 | 75 | 53% | 65% | 0% | – | – | 0.0 | 0.7 |
| SMP forehead (2 per person) | 160 | 80 | 40% | 80% | 3% | – | – | 0.0 | 0.2 |
| Diet study (strain-level) | 526 | 37 | 88% | 97% | 84% | 78% | 93% | 0.4 | 1.0 |
| Diet study (species) | 483 | 34 | 94% | 99% | 92% | 83% | 95% | 0.1 | 1.8 |
| MDSINE2 mice (healthy donor) | 320 | 5 | 3% | 48% | 0% | 0% | 32% | 0.1 | 13.2 |
| MDSINE2 mice (uc donor) | 381 | 5 | 0% | 16% | 0% | 0% | 18% | 0.0 | 15.7 |
| Diet study (strain-level) (2 per person) | 74 | 37 | 0% | 42% | 0% | – | – | 0.0 | 7.1 |
| IMP adults (16S) | 110 | 18 | 43% | 84% | 28% | 12% | 48% | 0.0 | 0.8 |
Student Microbiome Project (Flores et al. 2014, 16S, weekly), Immigration Microbiome Project (Vangay et al. 2018, 16S, monthly), ECAM (Bokulich et al. 2016, 16S, infants from birth), Minnesota diet study (Johnson et al. 2019, shallow shotgun, daily) and MDSINE2 mice (Gibson et al. 2025, 16S, every mouse given one donor's stool, a negative control). False flags are counted on samples that were left untouched.
Before you trust a flag
- Two or more samples per person. Three or more works better. With exactly two, SwapSeq can say which two people do not fit, but not always which tube moved.
- People who live together share microbes. Add a household column and look-alikes from one home move to the check list.
- Infants and same-donor animals carry little individual signal. SwapSeq says so with a low power estimate and rarely flags anything.
- Antibiotics and illness can make a sample unlike its owner without making it like anyone else. Those usually stay unflagged.
- A flag is a reason to check the lab records. Look at plate maps, extraction batches and sample sheets before you relabel or drop anything.
- Your data stays on your computer. Opening a file here reads it inside your browser. No server ever receives it.