Optimization and performance testing of a sequence processing pipeline applied to detection of nonindigenous species

Ryan Scott
Aibin Zhan
Emily A. Brown
Frederic J. J. Chain
Melania E. Cristescu
Robin Gras
Hugh J. MacIsaac, University of Windsor

Abstract

Genetic taxonomic assignment can be more sensitive than morphological taxonomic assignment, particularly for small, cryptic, or rare species. Sequence processing is essential to taxonomic assignment, but can also produce errors since optimal parameters are not known a priori. Here, we explored how sequence processing parameters influence taxonomic assignment of 18S sequences from bulk zooplankton samples produced by 454 pyrosequencing. We optimized a sequence processing pipeline for two common research goals: estimation of species richness and early detection of aquatic invasive species (AIS), and then tested most optimal models’ performances through simulations. We tested 1050 parameter sets on 18S sequences from 20 AIS to determine optimal parameters for each research goal. We tested optimized pipelines’ performances (detectability and sensitivity) by computationally inoculating sequences of 20 AIS into ten bulk zooplankton samples from ports across Canada. We found that optimal parameter selection generally depends on the research goal. However, regardless of research goal, we found that metazoan 18S sequences produced by 454 pyrosequencing should be trimmed to 375-400bp and sequence quality filtering should be relaxed (1.5 ≤ Maximum Expected Error ≤ 3.0, Phred Score = 10). Clustering and denoising were only viable for estimating species richness, because these processing steps made some species undetectable at low sequence abundances which would not be useful for early detection of AIS. With parameter sets optimized for early detection of AIS, 90% of AIS were detected with fewer than 11 target sequences, regardless of whether clustering or denoising were used. Despite developments in next-generation sequencing, sequence processing remains an important issue owing to difficulties in balancing false positive and false negative errors in metabarcoding data. This article is protected by copyright. All rights reserved.