What is SINTAX?
SINTAX (SImple Non-bayesian TAXonomy classifier) predicts the taxonomy of amplicon reads at all ranks — domain through genus — and attaches a bootstrap confidence value to each rank. It is implemented in the USEARCH sintax command and also in VSEARCH.
Key Design Choices
Unlike the RDP Naive Bayesian Classifier, SINTAX uses k-mer similarity (not Bayesian posteriors) to find the best-matching reference sequences. This means:
- No training step required
- Works with large, partially annotated databases (SILVA, Greengenes)
- Reference sequences don't need a lowest-rank label
- Comparable accuracy to RDP on V4 16S
- Lower over-classification on full-length sequences
Supported Markers
Bacteria & Archaea. All variable regions (V1–V9). Best results with full-length or V4.
Fungi. Use UNITE database. Species-level prediction sometimes possible.
Eukaryotes. Use SILVA eukaryotic subset.
Animated: query sequence broken into 8-mers → searched against reference database → top hits vote on taxonomy at each rank.
How SINTAX Works
Step 1 — k-mer word extraction
The query sequence is decomposed into all overlapping 8-mers. The same is done for all reference sequences in the database at index-build time.
8-mers: ACGTACGT, CGTACGTA, ...
Step 2 — Top-hit search
The fraction of shared 8-mers is used as a fast similarity metric to identify a shortlist of top-matching reference sequences from the database. This is the same k-mer word-counting approach used in USEARCH's global search.
────────────────────────
|kmers(Q) ∪ kmers(R)|
Step 3 — Taxonomy voting
The top-hitting reference sequences carry taxonomy annotations. SINTAX reads off the taxonomy of the single best-matching reference at each rank (domain → phylum → class → order → family → genus).
Step 4 — Bootstrap confidence
This is the key innovation of SINTAX. To assign a confidence value to each taxonomic rank, SINTAX repeats the k-mer search multiple times, each time using a random subset of the query's 8-mers:
Animated: 100 bootstrap replicates. Each replicate samples a subset of k-mers. The bootstrap confidence = fraction of replicates that agree on a given rank's prediction.
Build a k-mer index over all reference sequences at database creation time (makeudb_sintax). No training needed.
Find the best-matching reference using all query 8-mers. Extract its taxonomy string as the initial prediction.
Repeat the search 100 times, each time drawing a random subset (~¼) of the query's 8-mers. Record the top hit's taxonomy at each rank for every replicate.
confidence(rank) = fraction of bootstrap replicates that agree with the majority prediction at that rank. High-confidence predictions have values near 1.0.
Ranks below -sintax_cutoff (typically 0.8) are dropped from the final prediction. Output written to -tabbedout file.
Understanding Confidence Values
The bootstrap confidence is a reproducibility score, not a Bayesian posterior. It answers: "If I only used a subset of my sequence's k-mers, would I still get the same answer?"
# replicates predicting T_r
─────────────────────────
total replicates (100)
Confidence values are hierarchical: confidence decreases (or stays equal) as you go deeper in the tree. A genus prediction is always ≤ family confidence.
Interpreting the output
p:Firmicutes(0.99),
c:Bacilli(0.97),
o:Lactobacillales(0.93),
f:Lactobacillaceae(0.88),
g:Lactobacillus(0.76) ← below 0.8 cutoff → dropped
With -sintax_cutoff 0.8, genus is dropped from the final prediction since 0.76 < 0.8.
Animated bar chart: typical confidence decay through taxonomic ranks. The 0.8 threshold line is shown in orange.
SINTAX vs RDP Classifier
SINTAX was explicitly designed as a simpler alternative to the RDP Naive Bayesian Classifier (Wang et al. 2007). Both use bootstrapping for confidence, but differ fundamentally in the underlying method.
⚠️ Species prediction caveat
On short tags like V4, multiple species often share identical sequences. Bootstrapping cannot distinguish them — if the top hit is 95% identical, it likely belongs to a different species. Reference databases are also sparse (most microbial species are unnamed), leading to over-confident wrong species calls. Species-level predictions should be treated with extreme caution.
Usage & Parameters
Basic command
-db 16s.udb \
-tabbedout reads.sintax \
-strand both \
-sintax_cutoff 0.8
Build the database
-output silva.udb
Key parameters
| Parameter | Default | Notes |
|---|---|---|
-sintax_cutoff | none | Recommended: 0.8 |
-strand | — | Required: both or plus |
-db | — | FASTA or UDB with taxonomy |
-threads | 1 | Multithreading supported |
Recommended databases
🦠 16S (bacteria)
RDP training set v18 (21k seqs) — small, high quality, fast. SILVA v138 — larger, broader coverage but ~17% annotation errors.
🍄 ITS (fungi)
UNITE — current recommended database. Available from unite.ut.ee in SINTAX-compatible format.
🌿 18S (eukaryotes)
SILVA eukaryotic 18S subset (140k seqs). Broad coverage across protists, plants, fungi, animals.
⚠️ Database quality
~1 in 5 SILVA/Greengenes annotations are wrong (Edgar 2018). Small, authoritative databases (RDP, LTP) often give better genus accuracy despite lower coverage.