Towards predictive biomonitoring: integrating machine learning and ecological thresholds to identify potential...

Environmental DNA (eDNA) metabarcoding is increasingly used for monitoring anthropogenic impacts in marine ecosystems. Its applications include the identification of bioindicators and the development of eDNA-based biotic indices. In this study, we analyzed eDNA metabarcoding data of benthic foraminifera, which are known as sensitive to environmental changes, but their use in routine biomonitoring is limited by the difficulty of their morphological identification. We also used machine learning (random forest), complemented by Threshold Indicator Taxa ANalysis (TITAN) to identify potential bioindicators from foraminiferal metabarcoding data. Our study reveals clear shifts in foraminiferal diversity and community composition along distance from the fish cages, consistent with macrofauna-based ecological quality status classifications. We found that impacted sites near fish cages were dominated by some rotaliid species, while the organic-walled monothalamids were prevalent in less disturbed sites. Alpha and beta diversity metrics showed significant relationships with organic enrichment and other pollution-related parameters and benthic ecological indices. By combining machine learning-derived predictive importance with threshold-based ecological responses, several potential foraminiferal indicator ASVs were identified, some of them associated with known opportunistic taxa. Since benthic foraminiferal communities were strongly shaped by local environmental conditions and geographic location of farms, the candidate indicators require further validation across independent farms and regions. Overall, our results highlight the potential of integrating machine learning with ecological threshold analysis to identify candidate molecular indicators from eDNA metabarcoding data and provide a framework for developing genetic tools for aquaculture biomonitoring.