UNOISE3 & UCHIME2

Amplicon denoising and chimera detection algorithms powering modern microbiome analysis

16S rRNA ITS Illumina reads ASV denoising Chimera filtering
Denoising

UNOISE3 Algorithm

UNOISE3 performs error-correction (denoising) on amplicon reads, recovering correct biological sequences โ€” called ASVs โ€” and resolving sequences differing by even a single nucleotide.

Core Idea

For each unique read X with high abundance, UNOISE asks: are the lower-abundance neighbors of X just error-laden copies of X, or genuine biological sequences?

score(a, d) = ฮฒ(d) ร— ฮณ(a_ratio)
ฮฒ(d) = 0.5^(d ร— 1/ฮฑ)
ฮณ = a_X / a_Y

Where d = edit distance, a = abundance. Sequences with score < 1 are flagged as errors.

Two Sources of Bad Reads

1
PCR & sequencing errors

Point substitutions and indels introduced during amplification or sequencing. Usually low-abundance compared to parent.

2
Chimeric sequences

PCR artefacts formed by two parents joining mid-amplification. Handled by the integrated UCHIME3 step.

Animated: UNOISE3 neighbourhood model. Green = correct biological sequence; Red = error read; size = abundance. Hover over dots.

Algorithm Steps

1
Sort by abundance (descending)

Process most abundant unique reads first. Each becomes a candidate ASV.

2
Pairwise distance comparison

Compare each read to all confirmed ASVs using edit distance d. Check if score(a, d) โ‰ฅ 1 (genuine) or < 1 (error of known ASV).

3
Abundance threshold filter (minsize)

Reads below a minimum abundance (default: 2) are discarded as likely singleton errors regardless of score.

4
Chimera detection (UCHIME3 de novo)

Remaining sequences are checked for chimeric origin against the growing set of confirmed ASVs.

5
Output ZOTUs/ASVs

High-quality denoised sequences output. OTU table built separately with otutab.

โœ… Strengths

Single-nucleotide resolution. Resolves species/strains >97% similar. No arbitrary clustering threshold. Integrated chimera removal.

โš ๏ธ Limitations

Designed specifically for Illumina reads. Performance degrades on 454, Ion Torrent, or PacBio due to different error profiles.

โš™๏ธ Key Parameters

-minsize 2 โ€” minimum abundance
-unoise_alpha 2.0 โ€” controls sensitivity
-tabbedout โ€” per-read classifications

Chimera Detection

UCHIME2 Algorithm

UCHIME2 detects chimeric sequences โ€” PCR artefacts formed when a partial amplicon from one template anneals to a different template, producing a hybrid sequence. It extends the original UCHIME with improved scoring and filtering.

Animated: chimera formation during PCR. A partial amplicon from Parent A mis-primes onto Parent B.

The Chimera Score (H)

A candidate query Q is divided at a crossover point into left (L) and right (R) segments. Two parents P1 and P2 are found in the reference to best explain each half:

H = (id(Q, P1ยทP2) โˆ’ id(Q, best_parent))
/ penalty(ฮฒ, d)

If H exceeds a threshold, Q is flagged as chimeric. The score penalises low-divergence parents to reduce false positives.

Reference vs De Novo

R
Reference-based (uchime2_ref)

Parents searched from a large external database (SILVA for 16S, UNITE for ITS). Requires denoised inputs for low FP/FN.

D
De novo (uchime3_denovo)

Parents are the higher-abundance sequences within the same dataset. Integrated into UNOISE3 pipeline. Uses adjusted parameters to minimise false positives.

Algorithm Steps

1
Word-based parent search

Candidate parents for query Q identified by k-mer similarity search against the reference/database.

2
Alignment & scoring chunks

Q is divided into 4 chunks (by default). Each chunk is aligned to candidate parents. Voting identifies best left- and right-parent pair.

3
Score chimeric hypothesis

Compute H score: how much better is (P1 left + P2 right) vs best single parent? High H โ†’ chimera.

4
Classify: chimeric / borderline / non-chimeric

Three-way output allows downstream filtering. Borderline sequences carry uncertainty flags.

๐Ÿ“Œ Best Practice

Use uchime2_ref with the largest available database (SILVA/UNITE). Small "gold" databases miss many parents, causing poor sensitivity.

๐Ÿ”— UCHIME2 vs UCHIME3

UCHIME3 (uchime3_denovo) is the de novo variant built into UNOISE3, with tuned parameters to reduce false positives. UCHIME2 is recommended for reference-based mode.

โš™๏ธ Key Parameters

-uchimealns โ€” alignment output
-minh 0.28 โ€” minimum H score
-mindiv 0.8 โ€” min parent divergence

Algorithm Comparison

Choosing between the approaches depends on your data type and pipeline stage.

Feature UNOISE3 UCHIME2 (ref) UCHIME3 (de novo)
Primary purpose Denoising + chimera removal Chimera detection only Chimera detection (de novo)
Output ASVs / ZOTUs Chimera labels Chimera labels
Requires reference DB No Yes (SILVA, UNITE) No
Works on raw reads? Yes (Illumina) Only if denoised first Only if denoised first
Resolution Single-nucleotide โ€” โ€”
Best for 16S/ITS? โœ… Yes โš ๏ธ Only with denoised input โœ… Integrated in UNOISE3
Platform Illumina only Any Any (if denoised)

Recommended Pipeline

The USEARCH-recommended workflow for 16S/ITS amplicon analysis on Illumina data:

1. Quality filter

fastq_filter โ€” remove low-quality reads, truncate to fixed length

2. Dereplicate

fastx_uniques โ€” collapse identical reads, track abundances

3. Denoise

unoise3 โ€” error-correct + chimera filter โ†’ ZOTUs/ASVs

4. OTU table

otutab โ€” map original reads back to ZOTUs to get abundance table