Why the same stool sample gives different bacterial percentages on repeat consumer 16S / shotgun metagenomic tests
Every reproducibility question starts from the same thought experiment: take one bowel movement, divide it into several aliquots within seconds, and ask whether independent processing of each aliquot returns the same bacterial composition. It rarely does — and the gap between aliquots that should be identical is the baseline "noise floor" against which every consumer microbiome report must be judged.
A single stool specimen is not a homogeneous suspension of bacteria — it is a semi-solid matrix with microbial "hotspots," undigested fiber particles, mucus-associated communities, and luminal communities mixed unevenly. Studies swabbing multiple points of the same bowel movement before any homogenization step have found meaningful compositional differences purely from where the swab or scoop touched the specimen.
Most DTC kits ask users to swab a single point on toilet paper or take one scoop into a stabilizing tube (e.g., RNAlater-type or ethanol-based buffers, or OMNIgene-GUT). Because the tube is rarely vortexed to homogeneity before a lab draws its working aliquot, "the same sample" processed twice inside the same lab can already start from subtly different starting material — before extraction, PCR, or sequencing add their own variance.
The International Human Microbiome Standards (IHMS) project, published as Costea et al., "Towards standards for human fecal sample processing in metagenomic studies," Nature Biotechnology (2017), sent aliquots of the same homogenized stool samples to 15 laboratories across Europe, Asia, and North America and compared shotgun metagenomic results.
The finding that reshaped the field: the DNA extraction protocol used by a lab explained more of the variance between results than the actual biological differences between different people's microbiomes. In other words, two aliquots of the same stool, extracted with two different standard protocols, could look more different from each other than two different people's guts processed with the same protocol.
IHMS and follow-up inter-laboratory studies (Sinha et al., "Assessment of variation in microbial community amplicon sequencing by the Microbiome Quality Control (MBQC) project," Nature Biotechnology, 2017 — 15 laboratories, >20 protocol variants) both converged on the same conclusion: technical/protocol variance in stool microbiome sequencing routinely exceeds true inter-individual biological variance for several bacterial phyla.
The gut is an anaerobic environment; the moment stool leaves the body, it is exposed to atmospheric oxygen and ambient temperature. Bacteria keep growing, dying, and lysing during the hours between a bathroom collection and a freezer — and different taxa tolerate that window very differently, so the delay itself edits the result before any lab work begins.
Two independent processes run in parallel once a stool sample sits at room temperature:
• Differential die-off: strict (obligate) anaerobes such as Faecalibacterium prausnitzii and Roseburia spp. — both major short-chain fatty acid producers used as "gut health" markers in consumer reports — lose viability and their DNA begins to degrade once exposed to oxygen. • Differential overgrowth: facultative anaerobes, especially Enterobacteriaceae (E. coli and relatives, phylum Proteobacteria), tolerate oxygen and continue dividing at room temperature, so their relative share of the community rises the longer the sample waits.
Voigt et al., "Temporal and technical variability of human gut metagenomes," Genome Biology (2015), and later kit-comparison studies (Vogtmann et al., 2019 in PLOS ONE) both quantified this: samples left at room temperature for 24–72 hours before freezing showed significantly inflated Proteobacteria and reduced Firmicutes/Bacteroidetes ratios compared with samples frozen within 15 minutes.
Mailing a frozen sample is impractical for a consumer kit, so most DTC companies (uBiome historically, Viome, Thryve, DayTwo, Tiny Health) ship a collection tube pre-filled with a chemical stabilizing buffer — often ethanol-based, RNAlater-type reagents, or the DNA/RNA Shield and OMNIgene-GUT product families — that lyse and fix bacterial nucleic acids at the moment of collection, halting further community growth or die-off.
Stabilizer chemistry matters more than users assume: comparative studies find that different stabilizer chemistries preserve different taxa with different fidelity, and that samples collected in one stabilizer are not directly comparable in relative-abundance terms to samples collected in another. A customer who switches from one DTC brand to another between tests is therefore not just changing labs — they may be changing the chemistry that determines which bacteria survive to be counted.
A stool sample mailed in July heat versus one mailed in January cold, sitting in an uninsulated mailbox or postal sorting facility for 24–72 hours, can accumulate a comparable community shift to weeks of biological change in the gut itself — turning "ambient shipping time" into an uncontrolled experimental variable that most consumer reports never disclose.
Even a perfectly preserved sample must be broken open (lysed) to release DNA and then copied by PCR before it can be sequenced. Both steps are taxon-selective: some cell walls resist lysis, and some 16S primer/template pairings amplify more efficiently than others — so the final read counts reflect extraction and amplification chemistry as much as the original bacterial cell counts.
Gram-positive bacteria (many Firmicutes, including Clostridia clusters important for butyrate production) have thick peptidoglycan cell walls that resist chemical lysis alone; releasing their DNA efficiently generally requires mechanical bead-beating. Gram-negative bacteria (many Proteobacteria, Bacteroidetes) lyse more easily with gentler chemical methods.
A kit tuned for gentle, high-molecular-weight DNA yield (useful for long-read shotgun sequencing) will systematically under-represent tough-walled Gram-positive taxa relative to a bead-beating protocol optimized for total community recovery. The Human Microbiome Project standardized on validated bead-beating protocols after early comparisons showed extraction method alone could swing Firmicutes/Bacteroidetes ratio estimates by a factor of two or more in the same stool sample.
16S rRNA amplicon sequencing does not count bacterial cells directly — it counts copies of a marker gene. But bacteria carry between 1 and roughly 15 copies of the 16S rRNA gene per genome (catalogued in the rrnDB database), and copy number varies systematically by taxon: e.g., many Bacillus and E. coli strains carry 5–7 copies, while some Bacteroidetes carry closer to 1–4. A raw 16S read count therefore overestimates high-copy-number taxa and underestimates low-copy-number taxa relative to true cell abundance, unless a copy-number-correction step (e.g., PICRUSt2's normalization) is explicitly applied — which many consumer pipelines do not disclose.
On top of copy number, primers themselves bind different templates with different efficiency ("primer bias"). The V4 region (primers 515F/806R) and the V3–V4 region (primers 341F/805R) are the two most common amplicon targets in DTC testing, and published comparisons (e.g., Yang, Wang & Qian, "Reconstructing genetic maps with... 16S primer comparisons," and multiple V-region benchmarking studies) consistently show that switching only the target variable region — while sequencing the exact same DNA extract — changes the reported relative abundance of major phyla, sometimes reordering which phylum appears dominant.
Because copy-number and primer-binding biases are systematic rather than random, running the same DNA extract through two different 16S primer sets can produce two internally consistent but numerically different "microbiome reports" from a single tube of DNA — with no biological change in the gut at all.
After sequencing, the raw FASTQ file still has to be turned into a list of bacterial names and percentages — and that translation step is itself a major, underappreciated source of disagreement. Two labs can sequence the exact same DNA extract, use the exact same primers, and still report different species because they used different clustering algorithms, reference databases, or sequencing strategies entirely.
Older 16S pipelines clustered similar reads into Operational Taxonomic Units (OTUs) at a fixed similarity threshold, typically 97% — a coarse binning approach that treats reads within 3% divergence as "the same species." Modern tools (DADA2, Deblur) instead resolve exact Amplicon Sequence Variants (ASVs), distinguishing sequences that differ by a single nucleotide.
Re-analyzing the identical raw sequencing file with OTU clustering versus ASV inference routinely changes both the number of distinct "taxa" reported and their relative abundances, because ASV methods split what OTU methods lumped together (and apply different denoising/error-correction assumptions). A company that upgrades its bioinformatics pipeline between a customer's first and second test — without changing anything about the customer's gut — can report a different set of bacteria purely from the software update.
Taxonomic assignment requires matching a read against a reference database, and the major options — GreenGenes (largely unmaintained since ~2013), SILVA, RDP, and the newer Genome Taxonomy Database (GTDB) — do not use identical taxonomies. GTDB in particular has substantially reorganized bacterial phylogeny compared with legacy GreenGenes-based nomenclature (for example, reclassifying and renaming several long-used phylum and genus labels). A read that GreenGenes calls one genus can be assigned a different genus name — or a different phylum entirely — under GTDB, with no change in the underlying DNA sequence.
This matters directly for consumers: a "Firmicutes/Bacteroidetes ratio" reported by one company's legacy pipeline is not guaranteed to be numerically comparable to the same ratio reported by another company using a GTDB-based pipeline, even from identical raw sequence data.
16S rRNA amplicon sequencing reads one short, conserved marker gene and is inexpensive (roughly $100–150 per consumer sample) but typically resolves only to genus level, misses non-bacterial organisms (fungi, archaea, viruses) by design, and cannot detect functional genes (e.g., antibiotic resistance, metabolic pathways).
Shotgun metagenomic sequencing reads fragments of every genome present, typically at several million reads per sample, costs roughly $200–500+ per consumer sample, and can resolve species and even strain level while also profiling functional gene content — but requires far more input DNA, more sequencing depth, and more computationally intensive analysis (read alignment/assembly against much larger reference catalogs), introducing its own pipeline-dependent variability (e.g., Kraken2/Bracken vs MetaPhlAn vs HUMAnN give measurably different species lists from the same shotgun reads).
A consumer who takes a 16S test from one company and a shotgun test from another and expects the "percentage of Bacteroides" to match is comparing two different measurement instruments, not two readings of the same instrument.
Independent method-comparison studies running the same DNA through both 16S and shotgun pipelines have found percent-agreement on species-level composition dropping well below what marketing materials imply — dominant genera are usually confirmed by both methods, but strain-level and low-abundance taxon calls frequently diverge by an order of magnitude or disappear entirely between methods.
| Product | Indication | Trial Design | Key Result |
|---|---|---|---|
| 16S rRNA amplicon | Single marker gene (e.g., V4, V3–V4) | PCR amplify + sequence one variable region; OTU/ASV clustering | Cheap (~$100–150), fast, good for phylum/genus trends |
| Shotgun metagenomics | All DNA fragments present | Whole-community sequencing; classifier (Kraken2, MetaPhlAn) assigns reads | Species/strain resolution + functional gene content |
| OTU clustering (legacy) | 97% similarity bins | Groups similar reads into coarse taxonomic units | Simple, but merges closely related organisms |
| ASV inference (modern) | Exact sequence variants | DADA2/Deblur denoise to single-nucleotide resolution | Higher resolution, but not directly comparable to older OTU-based reports |
Stack every source of variance above — collection, storage, extraction, amplification, and bioinformatics pipeline — and the practical question becomes obvious: if a consumer mails two samples from the same week to the same or different DTC companies, how similar are the reports? Investigative testing and published literature both suggest the honest answer is "less similar than the marketing implies," and the regulatory framework covering these tests is thinner than most consumers assume.
Even with zero technical error, a person's gut microbiome is not static day to day: diet, sleep, stress, bowel transit time, and menstrual cycle phase all shift relative abundances. Longitudinal studies (e.g., the "Moving Pictures of the Human Microbiome" project, Caporaso et al., Genome Biology 2011, and later long-term cohorts) found meaningful compositional turnover even within the same individual across days to weeks — Bray-Curtis dissimilarity between two stool samples from the same healthy person taken days apart is routinely nontrivial, sometimes comparable in magnitude to the difference between two different individuals.
This means a consumer retesting a week later should expect some genuine biological drift on top of technical noise — but current DTC reports rarely distinguish "your gut actually changed" from "the assay just measured it differently this time," leaving users unable to tell the two apart.
uBiome, one of the earliest and best-funded consumer gut-microbiome testing companies (raised over $100M in venture funding), had its San Francisco offices raided by the FBI in April 2019 amid a federal investigation into insurance billing practices — the company had been billing insurers for microbiome tests framed as clinically necessary while marketing similar tests directly to consumers as wellness products. uBiome filed for Chapter 7 bankruptcy in September 2019, and in 2021 the U.S. Attorney's Office for the Northern District of California indicted uBiome's co-founders and former co-CEOs on fraud charges related to insurance billing.
The episode is instructive beyond the fraud allegations: it exposed how loosely regulated the boundary is between "clinical diagnostic test" (subject to CLIA, and in some cases FDA oversight) and "consumer wellness report" (subject mainly to general FTC advertising-truthfulness rules) — and how a single company could occupy both categories simultaneously with the same underlying assay.
Journalistic investigations (Vox, STAT News, and others, roughly 2018–2021) that split identical or near-identical stool samples across two or more DTC microbiome companies documented cases of meaningfully divergent "diversity scores," different dominant phyla, and contradictory dietary recommendations generated from the same person's gut within the same week — a real-world echo of the IHMS inter-laboratory findings from Stage 1.
Most DTC microbiome tests are run in laboratories certified under the Clinical Laboratory Improvement Amendments (CLIA, 1988) — but CLIA certification verifies that a lab follows sound general laboratory practices, not that a specific test's marketing claims ("optimize your gut score," "personalized probiotic match") are clinically validated. Many of these assays have historically been offered as Laboratory Developed Tests (LDTs), a category the FDA asserted authority to regulate more strictly in a final rule issued in April 2024 — though that rule faced significant industry litigation, underscoring how unsettled formal oversight of this space remains.
Several other laws that consumers assume apply often do not, cleanly: • HIPAA generally covers health plans, providers, and clearinghouses — a direct-to-consumer wellness company selling a gut test straight to an individual, outside insurance billing, frequently falls outside HIPAA's covered-entity definition, leaving state consumer-privacy laws (e.g., California's CCPA/CPRA) and FTC Act Section 5 (unfair or deceptive practices) as the main backstops. • GINA (Genetic Information Nondiscrimination Act, 2008) protects human genetic information from certain insurance/employment discrimination — but gut bacterial DNA is not human germline DNA, so microbiome results generally fall outside GINA's definition of "genetic information," a gap bioethicists have flagged as the field grows. • DSHEA (Dietary Supplement Health and Education Act, 1994) governs the "personalized probiotic" or "precision supplement" products several DTC microbiome companies sell alongside test results — supplements marketed under DSHEA are not FDA-reviewed for efficacy before sale, so a supplement recommendation generated from a noisy, non-reproducible test result is itself operating in a second layer of light-touch regulation.
Taken together, the technical reproducibility problem documented in Stages 1–4 exists inside a regulatory structure that was not originally built to police the clinical claims layered on top of it — which is why the burden of interpreting "why did my percentages change" still falls mostly on the consumer.