Predicting antifouling paint particle contamination based on 16S rRNA gene sequencing data using random forest-based machine learning

Antifouling paints often contain biocides designed to inhibit biological growth, and antifouling paint particles (APPs) have been previously shown to affect microbial communities in sediment. Given that typical methods for monitoring for APP presence can be specialized and challenging, alternative methods using simple, standardized, and universal approaches, such as 16S rRNA amplicon sequencing, would be highly valuable. This study uses a field-based mesocosm approach to train a random forest-based (supervised) machine learning model to predict APP presence and concentration in sediment based on 16S microbial community data. The model correctly predicted 100% of APP-presence samples and 83.3% of APP-absence samples in the incubation test set, although the model could not correctly predict APP concentration with sufficient accuracy. To determine the real-world applicability of the model, samples from 14 different sites along the Baltic Sea coastline and Warnow estuary in NE Germany were collected, and APP presence was pre-determined using scanning electron microscopy-energy-dispersive X-ray spectroscopy. The model correctly assigned APP-absence status to all APP-absent sites and correctly assigned three of five APP-contaminated sites as having APP presence. As such, these results confirm that it is possible to predict APPs in sediment based on the microbial community and serve as a proof of concept for the further development of machine learning-based predictive tools for environmental monitoring.IMPORTANCEThis study presents a groundbreaking pollution prediction model which is based on a new approach to training a machine learning model on microbial community data. A much lower amount of samples was needed using this novel targeted experimental approach to successfully train the pollution detection model. Most remarkably, when the model (which was trained on experimental samples from one site at one time of year) was applied to a variety of experimental samples from a range of geographical sites at a different time of year, the model worked, identifying all uncontaminated sites correctly and three of the five contaminated sites correctly, overcoming geographic and seasonal noise, which are typically major drivers for microbial community compositions. This work not only demonstrates that predicting antifouling paint particle presence in marine sediments using microbial community data is possible but also serves as a blueprint for the construction of other pollution detection models.