Machine learning analyses of bacterial oligonucleotide frequencies to assess the benthic impact of aquaculture

Abstract

Aquaculture is a rapidly expanding industry and is now one of the primary sources of all consumed seafood. Intensive aquaculture production is associated with organic enrichment, which occurs as organic material settles onto the seafloor, creating anoxic conditions which disrupt ecological processes. Bacteria are sensitive bioindicators of organic enrichment, and supervised classifiers using features derived from 16s rRNA gene sequences have shown potential to become useful in aquaculture environmental monitoring. Current taxonomy-based approaches, however, are time intensive and built upon emergent features which cannot easily be condensed into a monitoring pipeline. Here, we used a taxonomy-free approach to examine 16s rRNA gene sequences derived from flocculent matter underneath and in proximity to hard-bottom salmon aquaculture sites in Newfoundland, Canada. Tetranucleotide frequencies (k = 4) were tabulated from sample sequences and included as features in a machine learning pipeline using the random forest algorithm to predict 4 levels of benthic disturbance; resulting classifications were compared to those obtained using a published taxonomy-based approach. Our results show that k-mer count features can effectively be used to create highly accurate predictions of benthic disturbance and can resolve intermediate changes in seafloor condition. In addition, we present a robust assessment of model performance which accounts for the effect of randomness in model creation. This work outlines a flexible framework for environmental assessments at aquaculture sites that is highly reproducible and free of taxonomy-assignment bias.