Molecular and automated identification of the diatom genus Frustulia in northern Europe

Abstract

Molecular-assisted alpha taxonomy, which combines molecular species delimitation with post-hoc morphological examinations, has proved to be an effective tool for the classification of morphologically similar species. We employed this approach to examine the diversity of the genus Frustulia in northern Europe. First, we used two molecular markers to delimit species and then characterized their morphology using conventional and geometric morphometrics. Next, we employed machine-learning methods to identify valves in benthic diatom communities in order to infer the distribution and pH preference of individual species. Unlike previous studies using automated identification of diatom species, we examined the performance of the semi-supervised classifier. Supervised methods, which have been used before, only employ labelled valves to train the classification algorithm. The semi-supervised approach is, in addition, able to benefit from unlabelled valves in the natural populations. It is usually superior in cases in which there are few labelled data available. Finally, we compared the classification accuracy of the algorithms and five volunteer specialists. We found five molecular lineages, the F. crassinervia-saxonica species complex, F. gaertnerae, F. septentrionalis, F. krammeri, and F. cf. maoriana. The most valuable characteristics for species identification were length, width, striation pattern, and allometric shape changes described by the first axis in the geometric morphometric analysis. We found that a semi-supervised approach that does not rely solely on the morphology of isolated cells, but also accounts for variation among valves from natural populations, has superior performance. Based on valves from natural populations, we observed marked differences in species abundances and pH tolerances that have a bearing on their geographical distributions.