16S rRNA Amplicon Sequencing Explained: How a Read Becomes a Taxonomy Label

Quick answer: 16S amplicon sequencing (marker-gene profiling) reads one variable region of the 16S ribosomal RNA gene from every bacterium in a sample, then asks which known taxa the sequences match. The standard modern workflow is: trim primers, quality-filter and denoise reads (most often with DADA2), infer exact sequences called ASVs, count them into a feature table, and assign taxonomy with a trained classifier — most commonly inside QIIME 2. It is not the same as shotgun metagenomics, and its labels come with honest limits you should understand before quoting them.

Natural next step after the NGS workflow overview and the bioinformatics databases guide on this blog: same sequencers, same FASTQ files — the twist here is that every read comes from one conserved gene family.

What 16S sequencing actually measures

Every bacterium carries ribosomal RNA genes. The 16S rRNA gene is long enough (~1,500 bp) to contain conserved regions reliable “universal” primers can bind to, punctuated by variable regions (V1–V9) that differ enough between taxa to tell them apart. Sequencing one variable region (V3–V4 is the most common 2×300 chemistry choice) turns an entire microbial community into sequence data in one run — you do not have to culture anything.

That is the power and the boundary: you profile the marker gene, not the organisms’ other genes. Species-level calls through a single short region are often impossible because closely related species can share nearly identical 16S copies, and different strains carry different numbers of 16S copies to begin with.

Marker-gene profiling vs shotgun metagenomics

These get confused constantly, so pin the boundary early:

• 16S amplicon (marker-gene) sequencing targets one gene family. It answers “which taxa are here and roughly at what relative abundance?” cheaply, and works even when the species are unknown or the community is messy. No reference genome needed beyond the 16S database.
• Shotgun metagenomics sequences all DNA in the sample and maps it against genomes and genes. It can support functional answers (which pathways, which genes, which strains) and in-sample assembly, at a higher cost, higher wet-lab complexity, and heavier compute.

If your question is “which organisms are in this sample at what relative abundance,” a 16S marker-gene profile is usually the right budget-wise answer. If your question is “which genes enable that metabolic function,” you need shotgun. Mixed questions eventually benefit from both. And note the deeper limit both share: relative abundances are not an actual organism count — copy-number variation, primer bias, and sequencing depth distort them in known but imperfectly correctable ways.

The standard workflow, stage by stage

1. Demultiplex and inspect

Sequencing centers deliver samples already demultiplexed into per-sample FASTQ files plus a mapping file. Before anything else, look at the quality profiles: for V3–V4 2×300 runs, the tail ends of reads commonly fall below the quality where denoising still helps, which decides your truncation positions.

2. Dereplicate, denoise, and infer ASVs

With DADA2 (the two most common actions are filtering/trimming and denoising/pairing):

• Read pairs are trimmed to decided positions where quality keeps up.
• Errors are modeled per sample; algorithmically inferred “true” sequences are recovered.
• Read pairs from the same inferred exact sequence are merged.

The output is a set of ASVs — amplicon sequence variants — and a per-sample count table. An ASV is an exact sequence inferred to be a real biological unit rather than an error lump.

3. Which reference classifiers exist — and what “taxonomy assignment” really means

QIIME 2’s q2-feature-classifier ships several mechanisms, and “assign taxonomy” is not a single universal step. The main routes:

• A pretrained Naive Bayes classifier (sklearn-based), typically trained on full-width 16S references then applied to your region — the most commonly cited default.
• VSEARCH-based exact or near matches including consensus assignment.
• BLAST+-based alignment against reference databases of choice.

Each method, reference database (SILVA, GreenGenes2, RDP, UNITE for ITS), and region choice shapes results; label quality is bounded by database coverage, not only by the classifier.

4. Diversity, ordination, and the usual end products

Standard downstream questions proceed from the feature table and ASV sequences:

• Alpha diversity (within-sample richness and evenness per sample site or treatment).
• Beta diversity (between-sample dissimilarity via Jaccard or Bray-Curtis measures, usually visualized by PCoA or similar ordination).
• Differential abundance between conditions — usually the question that motivated the study — with statistical caveats about compositional data.

OTUs vs ASVs: what changed and what did not

The older clustering workflow (group sequences at ~97% similarity into OTUs) was built around databases and reproducibility-by-cluster. The modern denoising workflow instead infers ASVs as exact sequences — reproducible at the sequence level, comparable across studies down to the exact sequence, and no longer dependent on an arbitrary similarity threshold.

The honest framing is that the debate did not end with one method buried: ASVs are widely adopted as the better-resolved default, while OTUs persist in older studies and some ecological metric traditions. A practical consequence: cross-study comparisons work better at the sequence level; and counting “species” from a 16S table is already an overstatement — each ASV is an inferred sequence unit standing in for some taxa, not literally a species.

What a result can and cannot claim

A well-run 16S analysis legitimately supports claims of the form: “samples from condition X are enriched in sequences assigned to genus Y relative to condition Z.” It does not deliver:

• Definitive species-level identification (unless the region is unusually informative for your group).
• Functional inference from 16S alone is genuinely inference — pathway-abundance tools like PICRUSt-class tools predict functions from 16S profiles and are explicitly predictions from reference genomes, not measurements.
• Absolute abundance. Relative proportions shift when the whole community shrinks or sequencing depth changes; where needed, absolute counts require spike-ins or qPCR layering.

Common failure modes in practice

• Skipping quality inspection before choosing trim points. Read tails decide how long 16S fragments survive denoising.
• Mismatched primers or regions. A classifier trained on one region applied to another region degrades sensibly but silently; train on your actual amplicon region.
• Treating an ASV table as species censuses. Region informativeness plus copy numbers varies; treat labels as best-supported assignments, not identities.
• Blending 16S and shotgun claims. Taxonomy found by 16S and function found by shotgun are different evidentiary grades — cite each type appropriately.
• Comparing relative abundances without depth checks. Fewer total reads in a sample shuffles proportions with no biological cause.

Key takeaways

• 16S amplicon sequencing profiles one variable region of a conserved gene family — cheap, culture-free, taxonomy-first — and is not shotgun metagenomics.
• The modern pipeline: demultiplex, inspect, trim, denoise (DADA2 → ASVs), assign taxonomy with a trained classifier, then diversity and differential analyses.
• ASVs are exact inferred sequences; OTUs are approximate clusters at ~97%. Both appear in the literature; never call either literally species.
• Every label inherits database coverage constraints, plus copy-number and primer biases that relative abundances do not fully correct.
• Claims should be phrased to the evidence actually gathered: sequence units assigned to taxonomic ranks, not bacterial identities with functions.

Challenge

Download the QIIME 2 gut-to-soil beginner tutorial (the project’s current entry point for new users) and run its first feature-table step on a subset of samples, then answer: which steps produced your ASV table and how many ASVs per sample did you get, and what would change if the region were V1–V2 instead of V3–V4? Tight scope on one real run teaches the workflow better than any summary.

Need help planning an amplicon or metagenomics project for a course or lab? Ampersand Academy offers one-to-one mentoring in bioinformatics and analysis — part of Mahadhi, a development and digital marketing company.

—

FAQ

16S profiles one conserved gene family for a cheap taxonomy census. Shotgun sequences all DNA, supporting functional and strain-level answers at higher cost. They answer overlapping but distinct questions.

• What is the difference between 16S amplicon sequencing and shotgun metagenomics?

No. An ASV is an exact inferred sequence unit; multiple species or strains can share or differ within one ASV, and one species’ 16S copies can span more than one ASV. Treat labels as assignments with supporting evidence, not identities.

• Is an ASV the same as a species?

For sequence-level reproducibility and cross-study comparability, modern practice favors ASVs; OTUs remain in legacy studies. Choose deliberately, and cite unit definitions when comparing to older papers.

• Are ASVs better than OTUs?

SILVA, GreenGenes2, and RDP are standard for 16S; UNITE for ITS fungal work. Choose current, region-matched references, and record the exact training/reference version with your results.

• Which databases should I use for taxonomy assignment?

16S copy number varies by species and strain; primer and sequencing biases enter; reads are limited. Relative profiles approximate community structure, not cell counts. Convert to nearest absolute numbers only with external quantification such as qPCR or spike-ins.

• Why do 16S-derived relative abundances differ from actual organism counts?

Only as explicit prediction. Machine-learning tools that predict pathway abundances from 16S profiles exist, and they report what reference genomes suggest — not what was measured in the sample. Shotgun sequencing is needed for measurement.

• Can I infer function from 16S data alone?

No, but it is the most widely documented environment. DADA2 also runs directly in R, and equivalent pipelines exist in other tools. Any pipeline is fine when documented well; changing pipelines mid-study is what harms reproducibility.

• Do I need QIIME 2 to analyze 16S data?