A critical analysis of state-of-the-art metagenomics OTU clustering algorithms

Taxonomic profiling, using hyper-variable regions of 16S rRNA, is one of the important goals in metagenomics analysis. Operational taxonomic unit (OTU) clustering algorithms are the important tools to perform taxonomic profiling by grouping 16S rRNA sequence reads into OTU clusters. Presently various OTU clustering algorithms are available within different pipelines, even some pipelines have implemented more than one clustering algorithms, but there is less literature available for the relative performance and features of these algorithms. This makes the choice of using these methods unclear. In this study five current state-of-the-art OTU clustering algorithms (CDHIT, Mothur's Average Neighbour, SUMACLUST, Swarm, and UCLUST) have been comprehensively evaluated on the metagenomics sequencing data. It was found that in all the datasets, Mothur's average neighbour and Swarm created more number of OTU clusters. Based on normalized mutual information (NMI) and normalized information difference (NID), Swarm and Mothur's average neighbour showed better clustering qualities than others. But in terms of time complexity the greedy algorithms (SUMACLUST, CDHIT, and UCLUST) performed well. So there is a trade-off between quality and time, and it is necessary while analysing large size of 16S rRNA gene sequencing data.