Early marker-based metagenomic studies were performed without properly accounting for the effects of noise (sequencing errors, PCR single-base errors, and PCR chimeras). Denoising algorithms have been developed, but they were validated using data derived from mock communities, in which the true sequences were known. Since the algorithms were designed to be used in real community studies, it is important to evaluate the results in such cases. With this goal in mind, we processed a real 16S rRNA metagenomic dataset through five denoising pipelines. By reconstituting the sequence reads at each stage of the pipelines, we determined how the reads were being altered. In one denoising pipeline, AmpliconNoise, we found that the algorithm that was designed to remove pyrosequencing errors changed the reads in a manner inconsistent with the known spectrum of these errors, until one of the parameters was increased substantially from its default value. Additionally, because the longest read was picked as the representative for each cluster, sequences were added to the 3' ends of shorter reads that were often dissimilar from what had been removed by the truncations of the previous filtering step. In QIIME, the denoising algorithm caused a much larger number of changes to the reads unless the parameters were changed from their defaults. The denoising pipeline in mothur avoided some of these negative side-effects because of its strict default filtering criteria, but these criteria also greatly limited the sequence information produced at the end of the pipeline. We recommend that those using these denoising pipelines be cognizant of these issues and examine how their reads are being transformed by the denoising process as a component of their analysis.
