Neural networks and extreme gradient boosting predict multiple thresholds and trajectories of microbial biodiversity responses due to browning

Abstract

Ecological association studies often assume monotonicity such as between biodiversity and environmental properties although there is growing evidence that non-monotonic relations dominate in nature. Here we apply machine learning algorithms to reveal the non-monotonic association between microbial diversity and an anthropogenic induced large scale change, the browning of freshwaters, along a longitudinal gradient covering 70 boreal lakes in Scandinavia. Measures of bacterial richness and evenness (alpha diversity) showed non-monotonic trends in relation to environmental gradients, peaking at intermediate levels of browning. Depending on the statistical methods, variables indicative for browning could explain 5% of the variance in bacterial community composition (beta diversity) when applying standard methods assuming monotonic relations and up to 45 % with machine learning methods (i.e. extreme gradient boosting and feed-forward neural networks) taking non-monotonicity into account. This non-monotonicity observed at the community level was explained by the complex interchangeable nature of individual taxa responses as shown by a high degree of non-monotonic responses of individual bacterial sequence variants to browning. Furthermore, the non-monotonic models provide the position of thresholds and predict alternative bacterial diversity trajectories in boreal freshwater as a result of ongoing climate and land use changes, which in turn will affect entire ecosystem metabolism and likely greenhouse gas production.