Insights into the representativeness of biodiversity assessment in large reservoir through eDNA metabarcoding

mitochondrial reference database available in NCBI (https://ftp.ncbi.nlm.nih.gov/refseq/ release/mitochondrion/), and the nucleotide non-redundant (nr) database [36]. Only hits with an e-value lower than 1e-50 and a query cover higher than 90% were allowed. ASVs and MOTUs that matched off-target species were discarded. Off-target species included non- eukaryotic organisms, human, domesticated animals (Bos taurus, Canis familiaris, Felis catus, F. silvestris, Gallus gallus, Meleagris gallopavo, Mus musculus, Ovis vignei) and marine animals (Chimaera sp., Coryphaena equiselis, C. hippurus, Scomber sp., Homarus americanus, Rhineso- mus triqueter). Ichthyofauna communities For the pilot sampling, a Binomial Generalized Linear Model (GLM) was employed to estimate the probability of encountering fish species at each depth level, allowing for the assessment of the optimal sampling design. Heatmaps of the log10 frequencies of MOTUs were constructed based on MOTUs for fish species. MOTUs frequencies were computed considering all three markers (CO1, 16S, and 12S) combined. Based on fish MOTUs identified only up to the species level, ecological analyses of alpha- and beta-diversity were conducted in R. Two indices of alpha diversity were calculated with the vegan R [37] package within each locality: the Shannon Index (H’) [38], and richness (number of MOTUs). To compare the diversity indices across the 14 locations (beta-diversity), we constructed bar plot charts to visualize the species diversity in terms of fish orders. All

graphs were constructed in R scripts using the ggplot2 package [39]. Evaluation of the eDNA-based biodiversity inventory completeness To evaluate the completeness of our data, eDNA results were compared to the biodiversity data of TI built from Global Biodiversity Information Facility data (GBIF, https://www.gbif. org/) and a non-systematic literature review. In addition, both PCA and DAPC multivariate methods and rarefaction curve computation were employed. Using this integrative approach, we were able to assess the most informative samples and identify potentially redundant infor- mation. Only fish MOTUs identified up to the species level were considered, avoiding poten- tial bias in biodiversity assessments. The PCA was conducted to capture the variation among samples within the first two princi- pal components (PCs), considering both MOTUs abundance and richness per locality. Indi- vidual and scree plots were generated using the factoextra package [40] from R v. 3.3.2 [41], respectively. The DAPC was implemented to cluster samples based on compositional similar- ity. Because DAPC maximizes differences among groups, it enables the exploration of sample similarities and the identification of potentially redundant samples in the experimental design. The optimal number of PCs was determined by the optim.a.score function from the adegenet R package [42], and genetic clusters were identified using the find.cluster function. Individual cluster assignments from DAPC were visualized using the assignplot function from adegenet.

The contribution of each sample to the overall variability was assessed through computed eigenvalues within each PC. The impact of reducing sampling effort on capturing local variation was tested through the creation of four sub-datasets, three of them lacking one of the three most contributing samples and one containing only the three most contributing samples. Rarefaction curves were then computed using all complete and sub-datasets using the function specaccum from the vegan R package [43]. PLOS ONE eDNA metabarcoding reservoir PLOS ONE | https://doi.org/10.1371/journal.pone.0314210 January 24, 2025 6 / 22