Abstract
AbstractMuch effort has been made to understand why foliar microbes live where they do. However, whether foliar microbiome composition can be predicted is unknown. Here, we determine the limits of prediction using metabarcoding data of both fungal and bacterial assemblages that occur within (endophytes) and without (epiphytes) leaves from 59 plant taxa. We built random forest models for prevalent taxa and quantified the combined predictive power of 24 plant traits, 12 abiotic conditions and 7 additional features. As response variables, we considered microbial relative and absolute abundances, and occurrences. Most microbial taxa were too rare to effectively model, but model performance was generally poor even for the most prevalent and abundant taxa (model R2 was typically <0.1). Fungi were more tractable for modeling than bacteria. Models of Shannon’s diversity were moderately successful but those for richness were not. Taxa responded idiosyncratically and non-linearly to variation in the foliar habitat. When prevalent microbes were included as features in models, performance improved. Our results suggest that easily measurable aspects of the phyllosphere habitat are poor predictors of microbiome composition. These results pose a challenge for the study of microbial biogeography and we discuss possible ways forward.
Publisher
Cold Spring Harbor Laboratory