Roussel, J.-M.; Quemere, E.; Bonnet, B.; Covain, R.; Dezerald, O.; Lassalle, G.; Le Bail, P.-Y.; Petit, E. J.; Pottier, G.; Quartarollo, G.; Vigouroux, R.; Lalague, H.

1. Environmental DNA (eDNA) metabarcoding of water samples is increasingly used to detect fish species in streams. Several studies have concluded that it can outperform traditional inventory methods and recommend using it at large scales for fish-based ecological assessments. However, there is no standard protocol that can guarantee sufficient detection rates and repeatability, despite companies offering an extensive range of analyses. 2. We compared eDNA metabarcoding performed by four companies. Following their guidelines, samples were collected in a small tropical stream in the Maroni River (French Guiana) that hosts a species-rich fish community. We compared their inventories to each other and to a list of species captured during an extensive fish inventory performed immediately after sampling eDNA, as well as to current data on the species distributions. 3. The number of species detected by eDNA metabarcoding ranged from 5 to 48 among the companies, but these inventories contained many inaccuracies. All companies combined, 63 species were detected, of which 10 (16%) had never been reported in the Maroni River. The extensive inventory identified 50 species in the local fish community, of which 16-46 were not detected by eDNA metabarcoding (i.e. false negative detection rate of 32%-92% among the companies). 4. Reanalysis of raw sequencing data decreased differences among companies greatly, highlighting the importance of using a comprehensive and accurate DNA barcode database to assign species. Dissimilarity indices, calculated to compare the local fish community (based on presence/absence or fish catches) to eDNA detection, revealed large differences regardless of the company. 5. Synthesis and application. The large percentage of species not detected by eDNA metabarcoding of water samples could strongly bias fish-diversity inventories in streams that host species-rich communities. This issue is not well documented in the literature, and we recommend that similar studies in the future focus on other stream contexts. The large differences between commercial eDNA inventories and the local fish community challenge the use of eDNA metabarcoding for fish-based ecological assessments of streams. The variable performance of eDNA companies indicates the need for a standard protocol and access to a comprehensive DNA database before beginning large-scale eDNA programmes.