Reference Database Curation Shapes the Trade‐Off Between Predictive Performance and Ecological Interpretability in eDNA ‐Based Macroinvertebrate Bioassessment

ABSTRACT Environmental DNA (eDNA) is transforming global biodiversity monitoring and ecological assessment, yet linking eDNA signals to interpretable bioassessment information remains challenging, particularly when taxonomic assignments are compromised by incomplete or regionally biased reference databases. Here, we used morphological, eDNA, and physicochemical data from 233 sites across five watersheds in Zhejiang Province, China, together with a curated Chinese freshwater macroinvertebrate barcode library (CMBL), to assess how reference database choice affects SML prediction of chemical and biological water quality metrics. Morphological identification yielded 377 macroinvertebrate taxa. After quality control, the eDNA dataset contained 30,633 taxonomy‐free OTUs, of which 42.6% were assigned to prokaryotes or non‐target eukaryotes, 7.0% to the target taxonomic groups, and 50.4% remained unassigned. Under leave‐one‐out cross‐validation, taxonomy‐free models generally achieved the highest predictive performance, although non‐target sequences accounted for 58.5% and 43.0% of their top 200 features for WQI and BMWP prediction, respectively. However, performance declined under leave‐one‐watershed‐out cross‐validation, with R 2 values generally below 0.3, indicating limited transferability among watersheds. At high‐congruence sites, where eDNA and morphological family compositions closely corresponded, CMBL‐annotated models approached the performance of taxonomy‐free models for WQI prediction and slightly outperformed it for BMWP prediction. In contrast, predictive performance declined substantially at low‐congruence sites. These findings indicate that regional database curation can improve regional taxonomic relevance and family‐level congruence relative to a broad public reference database, although it does not guarantee reliable lower‐level taxonomic assignments. Our study highlights a trade‐off in eDNA‐based SML bioassessment: taxonomy‐free OTUs maximized predictive performance, whereas annotated OTUs provided taxonomic context whose reliability depended on reference quality, database coverage, and assignment confidence.