Machine learning identifies fish communities from environmental DNA (eDNA)

Biodiversity measurements of fish species are often conducted phenotypically, using extractive methods such as bottom trawls to physically capture and record organisms. In New Jersey waters, proposed offshore wind developments are amplifying the need for such measurements to to evaluate wind farm effects on fisheries. Environmental DNA (eDNA) metabarcoding offers a cost-effective, non-extractive method to survey fish biodiversity and abundance whilst minimizing environmental impacts. However, the complexity of high-dimensional eDNA data often hinders effective data analysis and clustering. We demonstrate that nonlinear dimensionality reduction techniques can outperform linear methods like PCA at visualizing high-level eDNA data in two dimensions. Following initial analysis and visualization of eDNA data, we utilized random forest machine learning (ML) models to predict fish species presence and community composition from oceanographic data. ML models were constructed for winter mixed, summer stratified, and combined seasonal data, using temperature and salinity as predictors. We found models were most accurate at making predictions in summer, correctly predicting species presence 80.6% and community cluster assignment 66.7% of the time. ML models also identified a significant contribution of oceanographic features towards fish community distributions in the Mid-Atlantic Bight. Future research can validate eDNA-derived community clusters against extractive trawl survey data to better inform fish stock assessments and regional offshore wind development.   

Presentation Preference: Poster

Primary Presenter: Henry Sun, Duke University (hs325@duke.edu)

Authors:

Henry Sun, Duke University  (hs325@duke.edu)

Josh Kohut, Rutgers University  (kohut@marine.rutgers.edu)

Jason Adolf, Monmouth University  (jadolf@monmouth.edu)