Machine learning-driven analysis of metabarcoding data to identify anthropogenic trace substances in aquatic ecosystems

Aquatic microbial communities play crucial roles in ecosystem functioning, providing a more comprehensive reflection of biological systems than single organisms. This makes them especially promising for evaluating ecosystem health. Although phytoplankton has been used in biomonitoring, bacterial communities offer additional insights. Despite their potential, bacterial communities and other microbial taxa, such as those revealed through 18S rRNA metabarcoding data, have not yet been fully integrated into official biomonitoring programs.

In this study, we demonstrated that using 16S and 18S rDNA metabarcoding data combined with machine learning enables the prediction of anthropogenic trace substances. Over one year, eDNA samples for 16S and 18S rRNA gene metabarcoding were collected twice weekly from 14 locations along the Warnow Estuary and Baltic Sea coast, with measurements of over 40 anthropogenic trace substances, such as pharmaceuticals, herbicides, and UV filters.

We trained Random Forest models to predict and quantify pollutants from the metabarcoding data, achieving high accuracy for a subgroup of contaminants. Using further regression analyses, we identified taxa that could be used as early warning indicators of substance concentrations exceeding ecotoxicological thresholds. Because the datasets used in this study span fresh water and marine habitats over the entire annual cycle, we expect the results to generalize easily to other locations.

Based on our findings, we propose that the integration of eDNA metabarcoding and machine learning will become a vital component of official biomonitoring programs, providing a scalable and efficient method for assessing ecosystem health and impacts of anthropogenic pressures in aquatic environments.