Limits to robustness and reproducibility in the demarcation of operational taxonomic units

The demarcation of operational taxonomic units (OTUs) from complex sequence data sets is a key step in contemporary studies of microbial ecology. However, as biologically motivated 'optimal' OTU-binning algorithms remain elusive, many conceptually distinct approaches continue to be used. Using a global data set of 887 870 bacterial 16S rRNA gene sequences, we objectively quantified biases introduced by several widely employed sequence clustering algorithms. We found that OTU-binning methods often provided surprisingly non-equivalent partitions of identical data sets, notably when clustering to the same nominal similarity thresholds; and we quantified the resulting impact on ecological data description for a well-defined human skin microbiome data set. We observed that some methods were very robust to varying clustering thresholds, while others were found to be highly susceptible even to slight threshold variations. Moreover, we comprehensively quantified the impact of the choice of 16S rRNA gene subregion, as well as of data set scope and context on algorithm performance. Our findings may contribute to an enhanced comparability of results across sequence-processing pipelines, and we arrive at recommendations towards higher levels of standardization in established workflows.