eDNA2IQI: an eDNA/random forest-based prediction tool that accurately predicts ecological status for regulatory-compliance assessment

Environmental monitoring is critical for assessing human impacts on ecosystems. In aquaculture, regulatory compliance has traditionally relied on biodiversity metrics derived from costly, time-consuming taxonomic identification of infauna. Environmental DNA (eDNA) metabarcoding offers a more efficient alternative; this study presents the development and validation of eDNA2IQI, an R package that processes bacterial-eDNA sequence data and, using machine learning models, predicts the associated infaunal quality index (IQI). The IQI is a key regulatory metric for monitoring aquaculture environmental impact under the EU's Water Framework Directive.

eDNA was extracted from 745 sediment grab samples from 75 aquaculture sites, with the traditional IQI calculated from macrobenthic identification. Using supervised machine learning (random forests) we optimised hyperparameters for 30 models using 5-fold cross-validation and blind testing on six full farm surveys. Model performance was analysed regarding prediction bias over the IQI range and various metadata categories. The eDNA2IQI R package was developed to standardise sequence data processing and IQI prediction, eliminating bias from any user input.

The results demonstrated strong agreement between predicted and traditional IQI values, with an R2 of 0.80, and an adjusted R2 of 0.95 in external validation on full site surveys, highlighting their effectiveness for regulatory monitoring. Models slightly overestimated low actual IQIs, and underestimated higher values, suggesting further optimisation is possible. There was no meaningful bias related to depth, sediment particle size, near-bed current speed or water temperature.

This study demonstrates the potential for eDNA-based methods combined with supervised machine learning to reduce both the cost and turnaround time of biodiversity impact monitoring.