How reliable is metabarcoding for pollen identification? An evaluation of different taxonomic assignment strategies by cross-validation

Metabarcoding is a powerful tool, increasingly used in many disciplines of environmental sciences. However, to assign a taxon to a DNA sequence, bioinformaticians need to choose between different strategies or parameter values and these choices sometimes seem rather arbitrary. In this work, we present a case study on ITS2 and rbcL databases used to identify pollen collected by bees in Belgium. We blasted a random sample of sequences from the reference database against the remainder of the database using different strategies and compared the known taxonomy with the predicted one. This in silico cross -validation (CV) approach proved to be an easy yet powerful way to (1) assess the relative accuracy of taxonomic predictions, (2) define rules to discard dubious taxonomic assignments and (3) provide a more objective basis to choose the best strategy. We obtained the best results with the best blast hit (best bit score) rather than by selecting the majority taxon from the top 10 hits. The predictions were further improved by favouring the most frequent taxon among those with tied best bit scores. We obtained better results with databases containing the full sequences available on NCBI rather than restricting the sequences to the region amplified by the primers chosen in our study. Leaked CV showed that when the true sequence is present in the database, blast might still struggle to match the right taxon at the species level, particularly with rbcL. Classical 10 -fold CV-where the true sequence is removed from the database-offers a different yet more realistic view of the true error rates. Taxonomic predictions with this approach worked well up to the genus level, particularly for ITS2 (5-7% of errors). Using a database containing only the local flora of Belgium did not improve the predictions up to the genus level for local species and made them worse for foreign species. At the species level, using a database containing exclusively local species improved the predictions for local species by similar to 12% but the error rate remained rather high: 25% for ITS2 and 42% for rbcL. Foreign species performed worse even when using a world database (59-79% of errors). We used classification trees and GLMs to model the % of errors vs. identity and consensus scores and determine appropriate thresholds below which the taxonomic assignment should be discarded. This resulted in a significant reduction in prediction errors, but at the cost of a much higher proportion of unassigned sequences. Despite this stringent filtering, at least 1/5 sequences deemed suitable for species -level identification ultimately proved to be misidentified. An examination of the variability in prediction accuracy between plant families showed that rbcL outperformed ITS2 for only two of the 27 families examined, and that the % correct species -level assignments were much better for some families (e.g. 95% for Sapindaceae) than for others (e.g. 35% for Salicaceae).