DADA2

DADA2 infers exact biological sequences at single-nucleotide resolution, producing Amplicon Sequence Variants as the fundamental unit of analysis. Unlike Operational Taxonomic Units (OTUs) that cluster sequences at an arbitrary 97% similarity threshold, ASVs resolve fine-grained biological variation. This enables:

• Distinguishing strain-level differences that OTU clustering obscures
• Reproducible labels across studies without reference to a clustering database
• Direct comparison of sequence variants across independent experiments

DADA2 constructs a data-driven error model from each sequencing run by alternating between sample inference and error rate estimation until convergence. The algorithm:

• Estimates substitution probabilities for each transition type (A→G, C→T, etc.) as a function of quality score
• Models Illumina-specific error profiles without relying on static, pre-computed error tables
• Adapts to run-to-run variability in sequencing chemistry and instrument performance
• Achieves error rates below 1 in 10,000 bases after correction

The core algorithm partitions dereplicated unique sequences into partitions consistent with the parametric error model. For each sequence abundance group:

• A p-value is computed testing the null hypothesis that a less-abundant sequence is an error of a more-abundant sequence
• Sequences are recursively partitioned until all partitions are internally consistent
• The most abundant sequence in each partition becomes the centroid ASV
• This approach avoids the greedy centroid selection bias present in heuristic denoisers like UNOISE

DADA2 performs model-based merging of overlapping paired-end reads, requiring a minimum overlap (default 12 bases) and exact concordance in the overlap region. The merger:

• Computes a consensus quality score for each position using the joint probability of agreement
• Rejects reads with mismatches in the overlap to eliminate residual errors
• Produces full-length denoised sequences spanning the entire amplicon
• Handles variable-length amplicons without requiring uniform read lengths

DADA2 identifies and removes PCR chimeras—artifactual sequences formed when incomplete extension products act as primers on heterologous templates. The method:

• Implements a de novo detection algorithm that identifies sequences explainable as combinations of two more-abundant parent sequences
• Uses a needleman-wunsch alignment to score potential parent-left and parent-right fragments
• Operates on a per-sample basis rather than pooling, preserving rare variants that might be flagged as chimeric in aggregate
• Removes bimeras—chimeras formed between two sample sequences—without requiring a reference database

DADA2's filtering step uses quality score profiles to truncate reads at the first position where quality drops below a threshold, rather than applying a uniform truncation length. The approach:

• Truncates forward and reverse reads independently based on their distinct error profiles
• Removes reads with ambiguous bases (Ns) or exceeding the maximum expected errors threshold
• Preserves maximum read length while eliminating low-quality tails
• Outputs a filtered quality profile for visual inspection of the truncation decision

The most common phylogenetic marker targeted in amplicon sequencing studies that use DADA2. This gene contains a mosaic of highly conserved and hypervariable regions (V1-V9). DADA2 is typically applied to amplicon pools targeting a single or multiple variable regions (e.g., V4). Key characteristics:

• Ubiquity: Present in all bacteria and archaea.
• Mosaic Structure: Conserved regions allow for universal primer design, while variable regions provide species-level phylogenetic signal.
• Reference Databases: ASVs are taxonomically classified against curated 16S databases like SILVA or Greengenes.

The legacy unit of analysis that ASVs supersede. OTUs are formed by clustering sequencing reads at a fixed similarity threshold, traditionally 97%. The limitations that DADA2's ASV method resolves include:

• Arbitrary Thresholds: The 97% cutoff is a heuristic that can lump distinct species together or split a single species into multiple OTUs.
• Non-Reproducibility: OTU clusters are defined relative to the specific set of sequences in a given study, preventing direct cross-study comparison.
• Loss of Resolution: True single-nucleotide biological variation is lost when sequences are collapsed into a consensus cluster centroid.

DADA2 implements both strategies to identify and remove chimeric sequences—artifactual PCR recombinants formed from two parent sequences. The choice depends on the study design:

• De Novo (Consensus): Identifies chimeras by comparing each sequence to more abundant "parents" within the sample itself. This is the default and is essential for discovering novel diversity not present in a reference database.
• Reference-Based (Remove Bimera Denovo): Compares sequences against a gold-standard reference database (e.g., SILVA). This method is more sensitive for known chimeras but cannot detect novel chimeric formations.