Microbial communities are commonly characterized by amplifying and sequencing target genes, but errors limit the precision of amplicon sequencing. We present DADA2, a software package that models and corrects amplicon errors. DADA2 identified more real variants than other methods in Illumina-sequenced mock communities, some differing by a single nucleotide, while outputting fewer spurious sequences. DADA2 analysis of vaginal samples revealed a diversity of Lactobacillus crispatus strains undetected by OTU methods.The importance of microbial communities to human and environmental health has motivated methods for their efficient characterization. The most common, and cost-effective, method is the amplification and sequencing of targeted genetic elements. Amplicon sequencing of taxonomic marker genes such as 16S rRNA [1], the ITS region [2] or 18S rRNA [3] provides a census of a community. Functional diversity can be probed by targeting functional genes [4].Disentangling errors from biological variation in amplicon sequencing data presents unique challenges, which has prompted the development of amplicon-specific error-correction methods [5,6,7,8]. Most of these methods were designed for pyrosequenced amplicons, and cannot be applied to Illumina sequencing.Currently, errors in Illumina-sequenced amplicon data are most often addressed by filtering low quality reads and constructing Operational Taxonomic Units (OTUs): clusters of sequences that differ by less than a fixed dissimilarity threshold (typically 3%) within which sequence variation is ignored [9,10,11]. Lumping similar sequences together reduces the rate at which errors are misinterpreted as biological variation, but OTUs under-utilize the quality of modern sequencing by precluding the possibility of resolving fine-scale (or strain-level ) variation [7,12,13,14,15]. Recent studies have shown that fine-scale variation can be informative about ecological niches [12,13], temporal dynamics [15], and population structure [4]. Fine-scale variation differentiates pathogenic from commensal strains in some cases [16,17], and can contain clinically relevant information for more complex microbiome-associated diseases [18,19,20].DADA -the Divisive Amplicon Denoising Algorithm -was introduced to correct pyrosequenced amplicon errors without constructing OTUs [7]. DADA was shown to identify real variation at the finest scales in 454-sequenced amplicon data while outputting few false positives [7,4].Here we present DADA2, an extension and reimplementation of DADA adapted for use with Illumina sequencing and available as an open-source R package available at https: //github.com/benjjneb/dada2. DADA2 implements a new model of Illumina-sequenced amplicon errors that incorporates quality information. Banded alignments and a kmerdistance screen improve computational performance. The DADA2 R package provides light-weight tools for other key parts of the amplicon denoising workflow: filtering, derepli-1 cation, chimera identification, and merging paired-end reads.We compared DAD...