Unravelling the outcome of 16S rDNA-based taxonomy analysis through mock data and simulations
Ali May, Sanne Abeln, Wim Crielaard, Jaap Heringa and Bernd W. Brandt
Department of Preventive Dentistry, Academic Centre for Dentistry Amsterdam (ACTA), University of Amsterdam and VU University Amsterdam, Amsterdam, The Netherlands.
Motivation: 16S rDNA pyrosequencing is a powerful approach that requires extensive usage of computational methods for delineating microbial compositions. Previously, it was shown that outcomes of studies relying on this approach vastly depend on the choice of pre-processing and clustering algorithms used. However, obtaining insights into the effects and accuracy of these algorithms is challenging due to difficulties in generating samples of known composition with high enough diversity. Here, we employ in silico microbial datasets to better understand how the experimental data are transformed into taxonomic clusters by computational methods.
Results: We were able to qualitatively replicate the raw experimental pyrosequencing data after rigorous adjustments on existing simulation software. This allowed us to simulate datasets of real-life complexity, which we used to assess the influence and performance of two widely used pre-processing methods along with eleven clustering algorithms. We show that the choice, order and mode of the pre-processing methods have a larger impact on the accuracy of the clustering pipeline than the clustering methods themselves. Without pre-processing, the difference between the performances of clustering methods is large. Depending on the clustering algorithm, the most optimal analysis pipeline resulted in significant underestimations of the expected number of clusters (min. 3.4%, max. 13.6%), allowing us to make quantitative estimations of the bacterial complexity of real microbiome samples.