Predicting hydrocarbon presence in marine cold seep sediments using machine learning models trained with benthic...

Hydrocarbon seepage in marine sediments exerts selective pressure on benthic microbiomes. Accordingly, microbial community composition in these sediments can reflect the presence of hydrocarbons, with specific groups being more prolific in association with seepage. Here, we tested machine learning models with large 16S rRNA gene amplicon data sets derived from marine sediments in deep-sea hydrocarbon prospective areas of the Eastern Gulf of Mexico and NW Atlantic Scotian Slope. Utilizing H2O's AutoML machine learning platform, it was determined that Gradient Boosting Machines performed best for creating 16S rRNA-based models that successfully predict the presence of hydrocarbons. Feature importance scores from the models revealed that in Gulf of Mexico samples, members of the Aminicenantia class (within the Acidobacteriota phylum) and the Sulfurovum genus (within the Campylobacterota phylum) were most diagnostic for the presence of low molecular weight hydrocarbon gases. The Campylobacterota lineage was also important in Scotian Slope sediments, along with sequences affiliated with the class-level JS1 group (within the Caldatribacteriota phylum) for determining hydrocarbon-positive sites. Testing these models in geographically distant seafloor basins showed that the microbial communities between basins varied sufficiently to prevent consistently accurate reciprocal predictions. However, models trained on a combined data set and filtered for important features performed substantially better, supporting the feasibility of generalized models under stringent feature selection. These results highlight the potential of seabed microbial taxonomy-based hydrocarbon seep site prediction when paired with refined sampling and consistent geochemical characterization.

Our study showcases an important use of bioinformatics in an interdisciplinary context, by combining hydrocarbon geochemistry and microbial biodiversity DNA sequencing profiles. We trained and compared different machine learning models on 16S rRNA-based bacterial taxonomy data using 377 DNA sequencing libraries from marine surface sediments in two different hydrocarbon prospective marine basins from different parts of the global ocean to predict the hydrocarbon status of sediment samples. Of all algorithms tested, Gradient Boosting Machines worked best for this objective. Feature importance scores from the models highlighted that in Gulf of Mexico samples, members of the Aminicenantales order and Campylobacterota lineages were most diagnostic for the presence of low molecular weight hydrocarbon gases. The Campylobacterota lineage was also important in NW Atlantic Scotian Slope sediments, along with sequences affiliated with the class-level group JS1 (within the Caldatribacteriota phylum) for determining hydrocarbon-positive sites, though several features appeared to be basin-specific. Importantly, models had a high prediction accuracy when predicting samples from the same basin but were less effective in predicting the hydrocarbon status in reciprocal basin testing, pointing to the ecological differences in hydrocarbon-driven environmental selection in different parts of the ocean. However, combined models using a refined set of predictive features improved cross-basin performance, highlighting the feasibility of a broader application. Results highlight the exciting potential of microbial taxonomy-based machine learning models in predicting broader ecological, oceanographic, and geological phenomena at the biosphere-geosphere interface.