16s Denoising Workflow

This guide explains the 16S denoising workflow, a method that resolves amplicon sequence variants (ASVs) from sequencing errors and replaces traditional OTU clustering. You should use this guide if you are a bioinformatician, microbiologist, or graduate student working with marker gene amplicon data and want a rigorous, practical framework for turning raw Illumina reads into high‑resolution ASV tables. The core decision is to shift from picking operational taxonomic units at a fixed similarity threshold to modeling and removing sequencing errors base by base, yielding single‑nucleotide resolution Galaxy Training Network: 16S Microbial Analysis. Two popular denoising engines , DADA2 and Deblur , each require careful parameter tuning, and this guide covers the critical decision points, step‑by‑step implementation, quality checks, and common pitfalls. Throughout, we anchor recommendations in evidence, using official training materials and peer‑reviewed method comparisons Bioconductor: DADA2 Pipeline.

At a Glance

Core Concepts and Decision Criteria

ASVs versus OTUs

Denoising replaces the older OTU clustering approach (typically at 97% identity) with exact sequence variants. ASVs have several advantages: they are reproducible across studies, provide higher taxonomic resolution, and avoid the arbitrary similarity cutoff. However, they require high‑quality input reads because each genuine variant must be distinguishable from a sequencing error. A systematic evaluation of amplicon workflows found that ASV‑based pipelines generally outperform OTU methods in detecting rare but real taxa, provided error modeling is robust Multi‑factorial examination of amplicon sequencing workflows.

Denoising algorithms

Two algorithms dominate the field. DADA2 (Divisive Amplicon Denoising Algorithm) learns an error model from the data and then applies a partition‑based correction. Deblur, integrated into QIIME 2, uses a Hamming‑distance filter after initial size filtering. The choice depends on your data characteristics. DADA2 tends to produce more ASVs and is more sensitive to parameter choices such as truncation length. Deblur is faster but may discard more low‑abundance sequences. For full‑length 16S (PacBio), newer tools like DADA2’s long‑read mode or dedicated clustering methods improve phylogenetic resolution De novo clustering of long‑read amplicons improves phylogenetic insight into microbiome data.

Primer trimming and read orientation

Primer removal is essential. If primers are not trimmed, denoising will treat them as part of the biological sequence, inflating error rates and creating spurious ASVs. Use cutadapt or QIIME 2’s cutadapt plugin. Decide whether primers were already removed by the sequencing center. For MiSeq v3 chemistry, reads often extend beyond the 2x300 bp limit, so you must consider overlapping regions. A common mistake is to assume primers were removed when they were not.

Quality filtering thresholds

The Phred score distribution dictates truncation points. Inspect per‑base quality plots using FastQC. For DADA2, a typical rule is to truncate reads where median quality drops below 30 (i.e., 1 error per 1000 bases). But the optimal point varies NCBI SRA: Quality Control Guidance. Over‑truncation shortens reads, reducing overlap and merging efficiency. Under‑truncation retains many errors, making error modeling unstable. Use the plotQualityProfile function in DADA2 to visualize the quality of your forward and reverse reads separately.

Workflow Implementation Steps

Step 1: Import and inspect raw reads

Obtain demultiplexed fastq files, typically from the sequencing facility or the NCBI Sequence Read Archive. Run FastQC on a representative subset to understand quality patterns, GC content, and adapter contamination. Use MultiQC to aggregate reports across samples. This step ensures you do not proceed with low‑quality or adapter‑laden data EMBL‑EBI Training: Quality Control of Sequencing Reads.

Step 2: Primer removal and quality filtering

Remove primers with cutadapt, allowing a mismatch rate of 0.1 to 0.2. Then apply quality filtering: for DADA2, use filterAndTrim() with truncQ=2, maxEE=c(2,2), truncLen values determined from quality plots. For Illumina reads, also set trimLeft to remove the first few bases that often have elevated error rates. The maxEE filter (maximum expected errors) is more lenient than a hard quality cutoff, it allows some lower‑quality bases as long as the sum of error probabilities stays below the threshold. Check the number of reads retained after each step. A loss of more than 40% of reads often indicates overly stringent parameters Galaxy Training Network: Quality Filtering with DADA2.

Step 3: Learn error rates

The denoising algorithm must estimate error rates from the data. In DADA2, learnErrors() fits a parametric model to the mismatch frequencies between reads and the inferred true sequences. Use at least 100 million bases for training (adjust by setting nbases). Convergence diagnostics: plot the error model using plotErrors(). The fitted lines (red) should follow the observed points (black) reasonably well, large systematic deviations indicate data problems or insufficient training Bioconductor: DADA2 Tutorial. If error rates look unrealistic (e.g., transition errors not higher than transversions), reconsider your truncation lengths or filter settings.

Step 4: Denoise and merge paired ends

Apply dada() to each sample or pooled across samples. Pooling improves sensitivity for rare variants but increases runtime. The algorithm outputs corrected reads and a table of sequence variants per sample. Then merge paired ends using mergePairs(). The minimum overlap should be at least 20 bp, with no mismatches allowed (or a maximum of 1 mismatch). If merging yields a low proportion of reads (under 60%), consider increasing truncation lengths to improve overlap or allowing more mismatches NCBI Bookshelf: Sequence Merging Guidelines. After merging, construct the ASV table with makeSequenceTable().

Step 5: Remove chimeras and construct final table

Chimera detection is critical because chimeric sequences are abundant in amplicon data. Use removeBimeraDenovo() in DADA2 with the method="consensus" to remove chimeras that appear in a majority of samples. Alternatively, use a reference‑based approach with a curated database. Check the fraction of reads lost to chimeras: typically 10‑20% is acceptable, higher loss may indicate PCR artifacts or library contamination. Export the clean ASV table as a BIOM file for downstream analysis.

Step 6: Assign taxonomy

Taxonomic assignment uses a reference classifier (e.g., the SILVA, Greengenes, or RDP databases). The assignTaxonomy() function in DADA2 uses a naive Bayesian classifier with a minimum bootstrap confidence of 80%. For higher resolution, use addSpecies() with a species‑level database, but be aware that species calls are less reliable for short amplicons (V4 region). The confidence level is a decision point: lowering it assigns more taxa but increases false positives. The Tourmaline workflow provides a containerized approach that standardizes taxonomy assignment and makes it iterable Tourmaline: A containerized workflow for rapid and iterable amplicon sequence analysis using QIIME 2 and Snakemake.

Step 7: Quality control of the final ASV table

After denoising, examine the total read count per sample, the number of ASVs, and the depth of discovery curves. Rarefaction curves (though not a normalization step) can reveal whether your sequencing depth captured the community diversity. Check for outliers: samples with very low read counts (below a few thousand) should be removed or re‑evaluated. Also inspect the distribution of ASV lengths: for a given hypervariable region, lengths should cluster tightly, outliers suggest artifacts or non‑target amplification.

Common Mistakes

One of the most frequent errors is ignoring per‑sample quality variations. Sequencing runs can have lane or cycle effects that cause some samples to degrade faster. Applying a uniform truncation length across all samples without checking individual profiles discards useful data from high‑quality samples or retains noisy data from poor ones. Always examine quality plots per sample rather than relying on an average.

Another mistake is failing to remove primers properly, especially for paired‑end data where primers may remain on one or both reads due to orientation. This leads to an inflated number of false ASVs because denoising algorithms interpret primer sequences as biological. Use cutadapt and confirm primer removal by checking that the first few bases in a sample match the expected biological sequence (e.g., the V4 region).

Over‑pooling during error rate learning can also be problematic. While pooling all samples improves sensitivity for rare variants, it can also mix very different error profiles (e.g., runs with different barcode lengths). A compromise is to pool samples from the same sequencing lane or to run learnErrors() on a subset of high‑quality samples only. The PCR cycle number and library preparation methods influence error rates, applying PCR cycle autonormalization to PacBio libraries showed that error rates are impacted by cycle counts Applying PCR cycle autonormalization to PacBio full‑length 16S rRNA library preparations. Although that study addresses long‑read data, the general principle holds for short‑read amplicons.

Finally, many users skip the step of inspecting the error model fit. If the error model does not converge, the denoising results may be unreliable. DADA2 provides diagnostic plots that should be examined before proceeding to merging and chimera removal.

Limits and Uncertainty

Denoising, while powerful, cannot correct systematic biases introduced during PCR amplification, such as primer bias or differential amplification of certain taxa. These biases persist in the final ASV table and can affect downstream analyses like differential abundance testing or co‑occurrence network inference Inferring microbial co‑occurrence networks from amplicon data: a systematic evaluation. The method also has limited ability to distinguish true low‑abundance ASVs from residual errors, very rare variants (with counts of one or two) should be treated with suspicion and often removed during prevalence filtering.

Another limitation is parameter sensitivity. Small changes in truncation length, maxEE, or chimera detection stringency can alter the final ASV set by tens to hundreds of variants. This is not necessarily a problem if the biological conclusions are robust, but it underscores the need for transparency in reporting parameters. Re‑analysis of the same dataset using different pipelines (e.g., DADA2 vs. Deblur) often yields different ASV tables, especially for the tail of rare variants. Researchers should be cautious when comparing absolute ASV counts across studies.

The biological interpretation is also limited by the taxonomic resolution of the reference database. Many environmental organisms have no close cultured relatives, so their taxonomic assignment may stop at family or genus level. A recent study on sorghum rhizobacteria showed that denoising produced finer resolution, but the authors still relied on functional inference to link ASVs to stress tolerance Exploring the Role of Rhizobacteria in Sorghum bicolor Adaptation to Combined Drought and Heat Stress. Denoising is a tool for generating high‑quality sequences, it does not replace careful experimental design or validation with complementary methods.

Frequently Asked Questions

Q: What is the difference between denoising and OTU clustering?
A: Denoising models sequencing errors at the base level to produce exact sequence variants (ASVs), whereas OTU clustering groups sequences at a fixed similarity threshold (often 97%). ASVs are reproducible across studies and can distinguish strains with single‑nucleotide differences. OTUs are more tolerant of sequencing error but lose resolution. For most modern amplicon studies, denoising is recommended as the primary approach Bioconductor DADA2 documentation.

Q: Can I use denoising for long‑read 16S data (PacBio, ONT)?
A: Yes, but with caution. Long‑read platforms have different error profiles (higher indel rates). DADA2 has a long‑read mode that works with PacBio CCS reads after quality filtering. A comparison of de novo clustering for long‑read amplicons showed that denoising improves phylogenetic insight when error rates are controlled De novo clustering of long‑read amplicons improves phylogenetic insight into microbiome data. You may need to modify truncation parameters and use a larger error learning set.

Q: How do I choose truncation lengths for DADA2?
A: Examine the per‑base quality plot for forward and reverse reads separately using plotQualityProfile. Set truncLen to a value where quality begins to drop sharply (e.g., median Q < 30). Avoid truncating too early because short reads reduce overlap and merging. A common strategy is to set forward truncation to 240 bases and reverse to 200 bases for V4 data, but always verify with your own plots Galaxy Training Network: DADA2 truncation.

Q: Should I rarefy my ASV table after denoising?
A: Rarefaction (subsampling to equal depth) is controversial because it discards data and reduces statistical power. Many modern differential abundance methods do not require rarefaction, they use count‑based models that account for library size (e.g., DESeq2, edgeR, ANCOM‑BC). For alpha diversity comparisons, rarefaction is still common but not mandatory. A safer approach is to use prevalence filtering and model‑based normalization. Always report the total reads per sample before any normalization.

References and Further Reading

• Galaxy Training Network: 16S Microbial Analysis with DADA2 and QIIME 2 , Step‑by‑step tutorials and workflows.
• Bioconductor: DADA2 Pipeline for Amplicon Data , Official R package documentation and case studies.
• EMBL‑EBI Training: Quality Control of Sequencing Reads , Practical guidance on fastq quality assessment.
• NCBI Bookshelf: Sequence Analysis and Quality Control , Authoritative technical references for bioinformatics.
• NCBI Sequence Read Archive: Quality Control Guidance , Information on raw read diagnostics.
• Tourmaline: A containerized workflow for rapid and iterable amplicon sequence analysis using QIIME 2 and Snakemake , Reproducible pipeline for denoising.
• Multi‑factorial examination of amplicon sequencing workflows from sample preparation to bioinformatic analysis , Systematic comparison of denoising versus OTU clustering.
• Applying PCR cycle autonormalization to PacBio full‑length 16S rRNA library preparations , Impact of PCR cycles on error rates.
• Inferring microbial co‑occurrence networks from amplicon data: a systematic evaluation , Explores limitations of denoising in network inference.
• Exploring the Role of Rhizobacteria in Sorghum bicolor Adaptation to Combined Drought and Heat Stress , Applied example of ASV‑based analysis.

Related Articles

• Fastq Read Quality Metrics
• Adapter Trimming Validation
• Read Duplication Sequencing
• Per Base Quality Sequencing
• Kmer Spectrum Analysis