The dada function now accepts a list of “priors”, i.e. sequences for which there is prior evidence they might be real. Input sequences that match one of the priors are evaluated against a relaxed threshold of statistical evidence (OMEGA_P instead of OMEGA_A), which allows those sequences to be detected at lower frequencies and even as singletons.

The dada function can perform “pseudo-pooling” with dada(..., pool="pseudo"). In pseudo-pooling, the input samples are denoised independently, then a set of sequences that appear in at least MIN_PREVALENCE samples are used as priors for a second and final round of sample inference. Pseudo-pooling approximates full pooling in linear time.

Error-correction can now be modified, and turned off, by the OMEGA_C parameter which sets the threshold at which error-containing reads are corrected (or not) to the sequence from which they are inferred to originate. OMEGA_C has been set to 1e-40 by default. In practice this has a very small impact on final abundances. The previous behavior (correct everything) can be recovered by setting OMEGA_C = 0.

seqComplexity calculates the complexity of input sequences, and can be used to identify and filter out low-complexity sequences.

plotQualityProfile now includes a cumulative description of read length variation.

The default minOverlap parameter of mergePairs was reduced from 20 to 12, and the alignment parameters used during merging were altered to more strongly penalize mismatches and gaps, which improves merging performance in repetitive sequences.

nbases has replaced the nreads parameter in the learnErrors function. As suggested by the name, this controls the amount of data the machine learning uses by the total number of bases rather than the read count, which is more appropriate given the range of read-lengths in target applications.

A new and extremely conservative form of greediness in the core denoising algorithm was added, providing some speedup in the core denoising algorithm. This heuristic can be toggled off by setting GREEDY=FALSE.

A new fast screen for optimal gapless alignments in the core denoising algorithm was added, providing some speedup in the core denoising algorithm. This heuristic can be toggled off by setting GAPLESS=FALSE.

The memory usage and speed of assignSpecies on large datasets has significantly improved.

Fixed an overflow bug on sequences 260nts or longer in the SSE=2 code. The DADA2 option enabling explicit 8-bit SSE vectorization in the C code was turned on by default (SSE=2). Some speedup in the core denoising algorithm.