The reference

SILVA

SILVA is a curated database of ribosomal RNA gene sequences — small subunit (16S/18S) and large subunit (23S/28S), across Bacteria, Archaea and Eukarya. Almost every 16S study uses it, and almost none of them state which of its three products they used or which release. Both matter.

Since 2023 SILVA has been hosted at the Leibniz Institute DSMZ, inside the DSMZ Digital Diversity consortium, which cross-links it with LPSN, BacDive and StrainInfo. Data from release 138 onward is CC-BY 4.0, and each release now carries a DOI — so there is no longer any excuse for "we used SILVA" as a methods sentence.

How a release is built

SILVA is not a dump of GenBank. It is a pipeline whose output is three databases of different stringency, plus a hand-maintained taxonomy.

INSDC / ENA — all rRNA gene sequences ragged, partial, some chimeric SINA → seed alignment columns are structurally homologous, not merely similar per-sequence scores alignment identityalignment quality Pintail (chimera anomaly)ambiguities · homopolymers SSU Parc alignment identity ≥ 50 · quality ≥ 40 everything that aligned at all SSU Ref ≥ 1,200 nt (900 for Archaea) · id ≥ 70 · qual ≥ 50 near-full-length, high confidence SSU Ref NR 99 Ref, dereplicated at 99% identity (VSEARCH) the one you classify against curated taxonomy — the part that changes between releases LPSN genus + phylum names, synchronised April 2023 six standard prokaryotic ranks · 43 validly published phyla uncultured replaced by Incertae Sedis · draws on GTDB, NCBI, UniEuk release 138.2 · DOI · CC-BY 4.0 ready-made classifiers: QIIME 2 (RESCRIPt / Clawback) DADA2 training sets · Kraken2

The three SSU products, side by side

ProductInclusion criteriaUse it forDo not use it for
Parc alignment identity ≥ 50, alignment quality ≥ 40 maximum-recall searching; finding anything remotely related taxonomic classification — it contains poor and partial sequences
Ref Parc criteria plus ≥ 1,200 nt (Bacteria, Eukarya) or ≥ 900 nt (Archaea), identity ≥ 70, quality ≥ 50 phylogenetic work needing near-full-length sequences; building custom references naive-Bayes training without dereplication — the redundancy skews the priors
Ref NR 99 Ref, clustered at 99% identity with VSEARCH amplicon classification. This is the default answer counting "how many sequences exist" — it is deliberately dereplicated

Releases, and the six-year freeze

Most people assume a reference database is a moving target they are perpetually behind on. For SILVA that has not been true for some time, and knowing it saves you from work you do not need to do.

ReleaseDateWhat it actually was
138Dec 2019Full release. Substantially revised prokaryotic taxonomy
138.1Aug 2020Taxonomy corrections to 138; LSU updated to match the SSU release
138.211 Jul 2024Maintenance only. LPSN adopted at genus and phylum level, a standard six-rank prokaryotic scheme, uncultured → Incertae Sedis
The sentence worth internalising

SILVA's own release notes describe 138.2 as "a maintenance release, with no additional sequences added to the ARB databases since the previous release 138.1." The counts did not merely stall — they went slightly down: SSU Parc −54, Ref −50, NR NR 99 −13, because 13 SSU and 7 LSU sequences with anomalously long branches were removed.

The sequence content of SILVA has not grown since August 2020. As of this writing there is no release 139, announced or otherwise.

Where that leaves you

9,469,070SSU Parc sequences
2,224,690SSU Ref sequences
510,495SSU Ref NR 99 sequences

If you built a classifier from SILVA 138 or 138.1, it is not out of date on sequence. What it is out of date on is taxonomy — 138.2 will give you different labels at genus and phylum for the same reads, and will say Incertae Sedis where the older release said uncultured. That is a relabelling, and it is the only reason to rebuild.

So the upgrade decision is unusually simple. Rebuild if you care about current nomenclature; otherwise pin what you have and say which release it is. Chasing SILVA releases is not a thing you need to do.

Using it against a short amplicon

SILVA sequences are near-full-length. Your reads are not. That mismatch is the single largest avoidable accuracy loss in the taxonomy step, and it is invisible — you get confident-looking labels either way.

16S rRNA gene · ~1,500 bases V1V2V3 V4V5V6 V7V8V9 515F / 806R → V4, ~253 bp everything else is unobserved trained on full-length SILVA reference query most of the trained signal is absent from the read taxa differing outside the window become inseparable extract-reads, then train reference query same window reference and query share a coordinate space silva-138-99-seqs-515-806.qza Genus is the honest floor for V4. Region-matched training improves accuracy and reduces overconfident deep calls. It does not make 253 bases separate congeneric species — nothing does. A species column in your taxonomy table is the reference's naming, not your evidence.

What the classifier is actually doing

The previous section says a region-matched classifier is better without saying what the classifier is. It is worth knowing, because the model explains both why it works and exactly which mistake breaks it.

training taxon A taxon B chop into overlapping 7-mers order is thrown away one 7-mer profile per taxon · hashed into 8,192 buckets taxon A profile taxon B profile prior · fit_prior: False → uniform a taxon is not favoured for having more references classifying one query read its own 7-mers multiply the word probabilities under each profile taxon A taxon B taxon B wins · confidence 0.62 threshold 0.70 → truncate to the deepest rank that clears it d__Bacteria; p__Bacteroidota; o__Bacteroidales; g__Prevotella; s__copri reported: …o__Bacteroidales — and no further

Three ways to put a name on a read

"Use BLAST" is not one method. The three in common use behave very differently, and only one of them is genuinely a bad idea.

MethodHow it decidesCan it say "I don't know"?Verdict
Best hit
top BLAST alignment		 Copies the label of the single closest reference No. Always returns a species, at any identity Avoid. Novelty becomes a confident wrong answer
Consensus alignment
classify-consensus-blast / -vsearch		 Takes the top N hits, votes, and backs off to the rank they agree on Yes — via min_consensus, and returns "Unassigned" when nothing clears perc_identity and query_cov Legitimate; competitive with naive Bayes when tuned
Naive Bayes
classify-sklearn		 Scores the read's 7-mer composition against a trained profile for every taxon Yes — truncates the label to the deepest rank clearing --p-confidence The QIIME 2 default: fast at inference, calibrated, tunable

Three properties separate the trained model from a plain best-hit search:

  1. It can abstain Best-hit has no mechanism for "this is a genus I have never seen". It returns the nearest neighbour's full species label whether the query is 99% or 85% identical. For most environments, most of your sample is not in any database — so this is not an edge case, it is the normal condition.
  2. It does not lean on a single identity number Percent identity is not calibrated across the tree: 97% 16S identity means different things in different clades, because 16S evolves at different rates and some genera are nearly identical at this locus while others are diverse. Any one global threshold over-splits some lineages and over-lumps others. A trained model learns clade-specific signatures instead.
  3. It pools the whole class Best-hit looks at one reference sequence and discards the rest. Naive Bayes builds its profile from every reference assigned to a taxon, so a read matching no single sequence especially well can still be placed by overall composition.
What the benchmark actually said

Bokulich et al. (2018) evaluated these on 19 mock communities and simulated novel sequences. The naive-Bayes, BLAST+-consensus and VSEARCH-consensus classifiers in QIIME 2 met or exceeded the species-level accuracy of other common methods — and the paper's emphasis is that parameter tuning drives performance. So naive Bayes is not strictly better than consensus alignment. It is the default because it is fast, calibrated and tunable, and because the alternative people reach for unaided is best-hit.

Its one real liability is the one this page opened with: it must be trained on your amplicon region. Alignment methods need no training, which is why a region mismatch hits them as mass Unassigned rather than as quiet degradation.

Practical rules

Pin the release, in the methods

"SILVA 138.2 SSU Ref NR99, classifier trained on the 515F/806R region" is a methods sentence. "SILVA" is not. The taxonomy changed substantially at 138 and again at 138.2 — phylum names, the six-rank scheme, uncultured → Incertae Sedis — so two studies on different releases are not directly comparable at those ranks.

Do not mix SILVA and GTDB names

They are different taxonomies with different rules, built for different purposes. SILVA curates rRNA-gene nomenclature against LPSN; GTDB reclassifies from genomes and normalises rank depth. Names collide without meaning the same thing. Pick one per analysis and say which.

Ref NR99 for classification

Parc is for searching, not labelling. Training a naive-Bayes classifier on undereplicated data lets thousands of near-identical submissions of well-studied organisms dominate the class priors — the dereplication in NR99 exists precisely to stop that.

Prefer a pre-built classifier when one matches

SILVA now publishes region- and habitat-specific QIIME 2 classifiers, DADA2 training sets and Kraken2 databases. If one exists for your primer pair, use it — it is reproducible, it is versioned, and training your own is an opportunity to get the orientation or the primer sequence wrong.

Read Incertae Sedis correctly

It replaced uncultured, and it means this lineage has no validly published placement at this rank — not unknown organism. A table dominated by Incertae Sedis at genus is telling you about the state of prokaryotic nomenclature in your habitat, not about a failure of your run.

Cite it

CC-BY 4.0 requires attribution, and each release now has a DOI, so cite the release you used rather than the original 2013 database paper alone. This is the cheapest reproducibility win available in the entire pipeline.

The failure this page exists to prevent

Taxonomy assignment fails quietly. A mismatched classifier does not error, does not warn, and returns a full table of plausible names. The only way to notice is to check something external — a mock community, a known control, or simply whether the phylum-level composition is credible for your sample type. If you have a mock community, classify it; if you do not, say so.

Alternatives, and what they believe

These are not interchangeable databases with different sizes. They embody different positions on what a taxonomy is for, and that is why their names collide without meaning the same thing.

ResourcePhilosophyConsequence in practice
SILVA Curate the rRNA gene itself; taxonomy follows validly published nomenclature (LPSN, Bergey's) Deep rRNA-specific curation and a structure-aware alignment. Bound to what has been formally named, hence the growing Incertae Sedis
GTDB Taxonomy should be derived from genomes and rank-normalised by evolutionary divergence Renames a great deal; ranks mean comparable things across the tree; releases roughly twice a year
Greengenes2 One phylogeny for 16S and shotgun data together, with taxonomy decorated onto it Lets amplicon and metagenomic results share a coordinate space; harmonised toward GTDB
RDP The original Bergey's-aligned curation; birthplace of the naive-Bayes classifier for 16S Effectively dormant as a database. Its training sets live on inside DADA2 and mothur
PR2 · UNITE · MIDAS Domain- or habitat-specific curation Eukaryotic 18S, fungal ITS, activated sludge. Inside their niche they beat any general database

The sharpest contrast is SILVA versus GTDB: nomenclature-first against genome-first. Pick one per analysis, say which, and do not mix the names.

Will it grow again?

A six-year freeze in sequence content is not an oversight. Several structural forces push the same way, and none of them is about compute.

  1. rRNA genes are systematically lost from assembled genomes The genome explosion does not feed SILVA. rRNA operons are repetitive and multi-copy, so assemblers collapse them and binners drop them. The million-plus MAGs that made MetaPhlAn 4 possible contribute very little full-length 16S. Genome data and rRNA data grow at completely different rates.
  2. The quality bar excludes almost everything being sequenced SSU Ref requires ≥ 1,200 nt, or ≥ 900 for Archaea. Essentially all 16S data generated today is 250–450 bp amplicons. Those land in Parc and never reach Ref — and Ref is what NR 99 is built from.
  3. NR 99 dereplicates the growth away A flood of new submissions from well-studied organisms adds nothing after clustering at 99%. Growth requires novel near-full-length diversity, which is the scarcest kind of sequence there is.
  4. Curation is manual and does not scale The taxonomy is hand-maintained against LPSN. Validly published names grow far more slowly than distinguishable lineages, which is precisely why Incertae Sedis keeps expanding.
  5. Sustainability The 2023 move to the Leibniz Institute DSMZ was framed explicitly as securing long-term sustainability. Institutions do not say that about resources with abundant funding.
The one counter-force

Full-length long-read 16S — PacBio and Nanopore — produces exactly the ≥ 1,200 nt sequences that SSU Ref requires, at scale, from environments no one has cultured. That is the plausible route to real growth, and it is why a 2026 survey on high-level taxonomy is more interesting than it sounds.

Practical upshot: do not expect SILVA to move under you. Pin your release and get on with the analysis. If you later need genome-anchored, rank-normalised taxonomy, or need to reconcile amplicon results with shotgun data, that is a reason to reach for GTDB or Greengenes2 — not a reason to wait for a newer SILVA.

Resources

WhatWhere
SILVA, current release documentationarb-silva.de/documentation/release-1382
SILVA news and release announcementsarb-silva.de/news
SILVA in 2026 — the DSMZ-era database paperNucleic Acids Research 54:D334
SILVA, the original database paperQuast et al., NAR 41:D590, 2013
QIIME 2 pre-formatted reference dataresources.qiime2.org
RESCRIPt — build and curate your own referencegithub.com/bokulich-lab/RESCRIPt
Classifier benchmarking and parameter tuningBokulich et al., Microbiome 6:90, 2018
fit-classifier-naive-bayes defaultsdocs.qiime2.org — fit-classifier-naive-bayes
classify-sklearn and --p-confidencedocs.qiime2.org — classify-sklearn
GTDBgtdb.ecogenomic.org
Greengenes2McDonald et al., Nature Biotechnology, 2023
Primer choice and region coverageKlindworth et al., NAR 41:e1, 2013

Every count and default on this page was read from the source linked beside it at the time of writing. Defaults move between releases — check --help in your own environment before quoting one in a methods section.