Uncovering microbial food webs using machine learning

Microbial trophic interactions are an important aspect of microbiomes in any ecosystem. They can reveal how microbial diversity modulates ecosystem functioning. However, uncovering microbial feeding interactions is a challenge because direct observation of predation is difficult with classical approaches such as behaviour and gut contents analyses. To overcome this issue, recent developments in trait-matching and machine-learning approaches are promising for successfully inferring microbial feeding links. Here, we tested the ability of six machine-learning algorithms for predicting microbial feeding links, based on species traits and taxonomy. By incorporating organism speed, size and abundance into the model predictions, we further estimated the probability of feeding links occurring. We found that the model trained with the boosted regression trees algorithm predicted feeding links between microbes best. Sensitivity analyses showed that feeding link predictions were robust against faulty predictors in the training set, and capable of predicting feeding links for empirical datasets containing up to 50% of new taxa. We cross-validated the feeding link predictions using an empirical dataset from a Sphagnum-dominated peatland with direct feeding observations for two dominant testate amoeba predators. The feeding habits of the two testate amoeba species were comparable between microscopic observations and model predictions. Machine learning thus offers a means to develop robust models for studying microbial food webs. It offers a route to combine traditional observations with DNA-based sampling strategies to upscale soil biodiversity research along ecological gradients.