UNOISE2: improved error-correction for Illumina 16S and ITS amplicon sequencing
Abstract
Amplicon sequencing of tags such as 16S and ITS ribosomal RNA is a popular method for investigating microbial populations. In such experiments, sequence errors caused by PCR and sequencing are difficult to distinguish from true biological variation. I describe UNOISE2, an updated version of the UNOISE algorithm for denoising (error-correcting) Illumina amplicon reads and show that it has comparable or better accuracy than DADA2.
02 /Specification & Classification
Click any tag to browse matching papers- Citations
- 1,818
- Organism
- Assay
- Research Focus
- Approach
In-depth summary
Claude Opus 5.5 · max effort · 8 Nov 2026 · from the full text of version 1 (15 Oct 2016)
Quality-score-free UNOISE2 clusters unique Illumina amplicon reads by abundance skew and edit distance with two preset parameters, matching or exceeding DADA2 accuracy on 16S mock communities.
Separating PCR and sequencing errors from real variants in 16S and ITS amplicon reads is hard, and DADA2 tackles it with quality scores and a per-dataset Poisson model of hundreds of parameters. UNOISE2 offers a one-pass, two-parameter alternative yielding zero-radius operational taxonomic units (ZOTUs) that keep apart strains 97% OTUs can merge, and intersecting its calls with DADA2's removes more errors than either alone.
Key results
ShownOn the Human Microbiome Project (HMP) Mock1 community, UNOISE2 returned no Noisy ZOTUs (near-matches likely carrying uncorrected errors) whereas almost half of DADA2's were Noisy; on the other mock datasets the two denoisers performed similarly.
Mock1: 22/50 DADA2 ZOTUs were Noisy (≥97% but <100% identity to a mock reference), versus 23/29 Perfect, 3/29 Exact and no Noisy ZOTUs from UNOISE2. On Mock2 most ZOTUs from both tools were Perfect or Exact, each tool giving ~100 Exact amplicons attributed to MiSeq cross-talk. · Validation; Tables
ShownOn the 27-strain Extreme mock community, UNOISE2 recovered more sequences absent from the design yet exactly matching SILVA (probable contaminants or unlisted variants) than DADA2, but missed one low-abundance designed strain that DADA2 detected.
All five non-reference, SILVA-exact ZOTUs were found by UNOISE2 and two of them by DADA2 (Table 3). UNOISE2 gave two designed strains one Perfect plus one Noisy ZOTU each; all other designed-strain ZOTUs were Perfect for both tools. UNOISE2 missed P. copri, which DADA2 recovered. · Validation; Tables
ShownKeeping only ZOTUs called by both UNOISE2 and DADA2 (the U&D intersection) preserved most Perfect and Exact sequences while removing most Noisy near-matches, consistent with the two algorithms making different mistakes.
In all compared datasets, U&D kept most Perfect (100% to a mock sequence) and Exact (100% to SILVA v123) ZOTUs while dropping most Noisy and many Good (>97% to a database sequence) and Other ZOTUs; exact counts are given only in Table 2, not the text body. · Introduction; Validation; Tables
ShownOn the high-diversity Soil1 16S sample, DADA2's removeBimeraDenovo labeled over twice as many DADA2 amplicons chimeric as UCHIME2 (the filters agreed fully on low-diversity mock data), which largely explains UNOISE2's higher one-difference neighbor count (N1) there.
With both filters applied to DADA2 amplicons, DADA2 flagged 939 of 8,470 versus 401 for UCHIME2, all but one of which DADA2 also flagged. Of the 539 DADA2-only calls, 282 had parent/query abundance ratios between one and two and 257 used a chimeric model with one difference. · Validation; Chimera filtering; Figure legends
SuggestedThe abundance-ratio profile of DADA2-only chimera calls conflicts with PCR expectations (parents of a true chimera should be at least roughly twice as abundant), so the author infers most are false positives from coincidental parent models, and possibly some UCHIME2 calls too.
With 30 PCR cycles, the author's upper-bound estimate is ~2% (~1/60th) of true chimeras having ratio <2, yet 282/539=52% of DADA2-only calls had ratio <2 and 90/539=17% had ratio <1.2, with counts rising as the ratio fell toward one; UCHIME2 matched 400/657 DADA2 calls with ratio >2. · Chimera filtering; Figure legends
Methods, data and reproducibility
Unique sequences with abundance ≥γ (default 4) are scanned by decreasing abundance; each joins a centroid if its abundance skew is ≤β(d), with β decaying exponentially in Levenshtein distance d (α=2: β(1)=1/8), else founds a new one. Centroids, chimera-filtered by UCHIME2, become ZOTUs; outputs were benchmarked against DADA2 on mock and in vivo MiSeq data.
- Mock1: Human Microbiome Project (HMP) mock community, MiSeq paired-end 16S V4 reads; its most abundant unique sequence is 233 nt after pair merging and primer trimming · 21-strain mock incl. Staphylococcus epidermidis and S. aureus · n = 21 strains
- Mock2: second HMP-designed mock community MiSeq paired-end run, randomly subsampled because DADA2 did not converge on all reads · 21-strain mock community · n = 1M read pairs (random subset)
- Extreme mock community dataset previously used to validate DADA2 (MiSeq paired-end 16S reads) · 27-strain mock community incl. P. copri · n = 27 strains
- Soil1: high-diversity in vivo soil 16S sample amplified with 30 PCR cycles (MiSeq paired-end), randomly subsampled · n = 1M read pairs (random subset)
- In vivo vaginal sample MiSeq paired-end reads, denoised by both tools (DADA2 emitted 187 sequences twice, differing only in length)
- In vivo ITS amplicon dataset (MiSeq paired-end), denoised by UNOISE2 only because DADA2 was not run on ITS
- HMP mock community 16S reference sequence database used to classify ZOTUs as Perfect or Noisy · 21-strain mock community · n = 115 sequences for 21 strains
- Several unnamed mock and in vivo Illumina datasets whose FASTX-LEARN error abundance profiles guided the hand-designed β function
- Methods & tools
- UNOISE2 (abundance-skew denoiser; defaults α=2, γ=4), DADA2 (comparator Illumina denoiser), DADA2 removeBimeraDenovo chimera filter, UCHIME2 (denoised de novo chimera filtering), FASTX-LEARN (error abundance distribution profiling), UCLUST (97% identity clustering of ZOTUs into OTUs), SILVA 16S database v123, UNITE ITS database v04.11.2015, Expected-error read filtering (discard reads with ≥1 expected errors), Bayesian paired-read assembler (recommended merging step), Illumina MiSeq paired-end sequencing
- Software & environment
- Only reference database versions are given: SILVA 16S v123 and UNITE ITS v04.11.2015; software versions for UNOISE2, DADA2, UCHIME2 and UCLUST are not stated.
- Reproducibility gaps
- Only link is a FASTX-LEARN manual page; no UNOISE2 code repository, dataset accessions or DADA2 run settings are given, and Mock2 and Soil1 were randomly subsampled to 1M read pairs with no seed reported.
Limitations (6)
- The β skew-threshold function is a hand-built phenomenological model with no underlying error theory, and the author states it fits observed error abundance distributions only very roughly. · stated by the authors
- Conservative thresholds favor lumping and cost sensitivity: correct uniques below γ=4 are lost, some d=1 or d=2 variants merge into one ZOTU, and pooling can mistake a sample-specific close variant for an error. · stated by the authors
- Ground truth is uncertain: the HMP mock reference (21 strains) lists 115 distinct 16S sequences, Noisy ZOTUs may be real variants (paralogs, impure cell lines), and no gold standard or realistic simulation can score chimera filtering because low-divergence chimeras mimic point errors. · stated by the authors
- The conclusion that most DADA2-only chimera calls are false positives rests on assumptions and approximations the author concedes are hard to test; UCHIME2 may instead have missed valid chimeras, which would call for running it with DADA2-equivalent criteria. · stated by the authors
- Benchmark scope is narrow: three mock and three in vivo MiSeq datasets with no replicate-level statistics, no ITS head-to-head (DADA2 not run on ITS), and Mock2 and Soil1 subsampled to 1M read pairs after DADA2 failed to converge, leaving full-depth performance untested. · noted in this summary
- The β function (default α=2) was hand-fitted to error profiles from several unnamed mock and in vivo Illumina datasets whose independence from the validation sets is not stated, so the benchmark may favor UNOISE2. · noted in this summary
Who should read:Read this if you denoise 16S or ITS amplicons and are weighing UNOISE2 against DADA2, or want an abundance-ratio argument that permissive de novo chimera criteria yield many false positives.
Reproducibility ledger
Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.
Sign in to file a reproducibility report.
Related papers
- AmpliCI: A High-resolution Model-Based Approach for Denoising Illumina Amplicon Data
Xiyu Peng, Karin S. Dorman·BioinformaticsPeer-reviewed·Jul 16, 2020
- POAnoise: A Graph-based Denoising Pipeline for Amplicon Sequencing Data
Muhammad Ardiyansyah, Brendan Furneaux, Otso Ovaskainen·bioRxiv·Sep 21, 2026
- MAUI-seq: Metabarcoding using amplicons with unique molecular identifiers to improve error correction
B. Fields, S. Moeskjaer, … J. P. W. Young·bioRxiv·Jan 12, 2020
- NanoAmpli-Seq: A workflow for amplicon sequencing for mixed microbial communities on the nanopore sequencing platform
S. T. Calus, U. Z. Ijaz, A. J. Pinto·GigaSciencePeer-reviewed·Nov 23, 2018
- DADA2: High resolution sample inference from amplicon data
Benjamin J Callahan, Paul J McMurdie, … Susan P Holmes·bioRxiv·Aug 6, 2015
- Long-read amplicon denoising
Venkatesh Kumar, Thomas Vollbrecht, … Ben Murrell·Nucleic Acids ResearchPeer-reviewed·Jul 24, 2019