From GenBank to GBIF: Phylogeny-Based Predictive Niche Modeling Tests Accuracy of Taxonomic Identifications in Large Occurrence Data Repositories

Abstract

Accuracy of taxonomic identifications is crucial to data quality in online repositories of species occurrence data, such as the Global Biodiversity Information Facility (GBIF), which have accumulated several hundred million records over the past 15 years. These data serve as basis for large scale analyses of macroecological and biogeographic patterns and to document environmental changes over time. However, taxonomic identifications are often unreliable, especially for non-vascular plants and fungi including lichens, which may lack critical revisions of voucher specimens. Due to the scale of the problem, restudy of millions of collections is unrealistic and other strategies are needed. Here we propose to use verified, georeferenced occurrence data of a given species to apply predictive niche modeling that can then be used to evaluate unverified occurrences of that species. Selecting the charismatic lichen fungus, Usnea longissima, as a case study, we used georeferenced occurrence records based on sequenced specimens to model its predicted niche. Our results suggest that the target species is largely restricted to a narrow range of boreal and temperate forest in the Northern Hemisphere and that occurrence records in GBIF from tropical regions and the Southern Hemisphere do not represent this taxon, a prediction tested by comparison with taxonomic revisions of Usnea for these regions. As a novel approach, we employed Principal Component Analysis on the environmental grid data used for predictive modeling to visualize potential ecogeographical barriers for the target species; we found that tropical regions conform a strong barrier, explaining why potential niches in the Southern Hemisphere were not colonized by Usnea longissima and instead by morphologically similar species. This approach is an example of how data from two of the most important biodiversity repositories, GenBank and GBIF, can be effectively combined to remotely address the problem of inaccuracy of taxonomic identifications in occurrence data repositories and to provide a filtering mechanism which can considerably reduce the number of voucher specimens that need critical revision, in this case from 4,672 to about 100.

The Global Museum: natural history collections and the future of evolutionary science and public education

“…As the main global biodiversity database, a large proportion of the >1 billion records comprise observations rather than specimens (see below). Smith, Johnston & Lücking (2016) and Yesson et al (2007) discuss issues regarding data quality in GBIF, such as unreliable taxonomic identifications in the absence of voucher specimens, and non-global coverage of species distribution data. Verifying suspect occurrences through niche modelling, based on verified and geo-and DNA-referenced occurrences of the same species, is a step toward identifying unreliable records.…”

Climate Change, Bioclimatic Models and the Risk to Lichen Diversity

“…However, bioclimatic modelling of lichens has been used for a wide variety of purposes, and many such studies have explored the relationship of lichen distribution to baseline climate only, alongside a range of other covariables, and without projection to climate change scenarios (Table 1). Thus, lichen bioclimatic modelling has been used to test taxonomic hypotheses [32][33][34][35][36], improve understanding of threatened species [37][38][39][40][41][42][43][44][45] including through conservation design [46,47], identify indicator species [48][49][50][51], or to test and improve the practical application of bioclimatic methods [37,[52][53][54]. These studies at the baseline highlight three key interrelated decisions characterising the development process for any bioclimatic model: 1.…”

“…Lichen distributions are dynamic [55] and any shifts within this baseline period are discounted; there is an important balance to be struck between a time period that extends to provide a reliable distribution, minimising issues of spatial bias [56,57], while constraining this period to represent as stable a distribution as is possible with respect to prevailing climate. Furthermore, field occurrence records provide 'presence-only' data which present an additional statistical challenge that has been handled along a continuum, as follows: (i) by generating a constrained set of pseudo-absences [58,59] for use with standard forms of regression such as generalised linear or additive models [33,39] or with alternative methods that facilitate nested interactions such as classification and regression trees including random forest [33], (ii) using a controlled selection of 'background' pseudo-absence points as applied in MAXENT [34,35,40,[43][44][45], or alternatively (iii) using presence-only statistical methods that compare occurrences to the properties of an entire environmental 'background' [38,46,47]. Lichen bioclimatic models have thus used a rich variety of statistical techniques (Table 1), extending to include nonparametric multiplicative regression that has been applied to >50% of studies with abundance or presence-absence data [36,48,50,51].…”

“…Lichen bioclimatic models at the baseline have also varied in their extent, from global [34] to landscape-scale [41]. Alongside extent, model resolution has been determined by a match between the scale used to generate the lichen response and that of available climate predictors.…”

Advancing DNA Barcoding and Metabarcoding Applications for Plants Requires Systematic Analysis of Herbarium Collections—An Australian Perspective

“…The responsibility for accurate initial specimen identification and sequence data quality lies solely with the data generator for standard submissions to GenBank and there is often very little information upon which veracity can be assessed (e.g., voucher specimen meta-data, raw sequencing data files). Reliance on data that has not been derived from well curated and vouchered specimens poses risks of misidentification, an issue that is well recognized for GenBank data (Crocetta et al, 2015;Smith et al, 2016;Balakirev et al, 2017). A recent study of fish barcode sequences found evidence for potential errors in ∼4% of sequences (Li et al, 2018).…”