Local microbial yield-associating signatures largely extend to global differences in plant growth

Rapid advancements in high-throughput DNA sequencing have opened new avenues for applying microbiome-based machine learning to predict and model determinants that enhance agricultural productivity and sustainability in agroecosystems. Although early attempts have been made to predict crop yield or measures of soil health through the soil microbiome, it is unclear if microbial patterns associated with plant growth or crop yield on a local scale can be generalized to predict differences in plant growth on a continental or global scale. Herein, we measured the soil bacterial microbiome on a single maize field in Germany with high spatial sampling resolution and correlated the community composition with corresponding volume flow-based high-resolution yield measurements. Applying machine learning techniques, a least absolute shrinkage and selection operator (LASSO) regression model could retrospectively predict ∼65 % of variation in maize yield through cross-validation. We validated this locally trained model, comprising 26 genera, using data from seven publicly available datasets. Predictions from this model correlated with various yield or plant growth metrics throughout the world and could predict up to 37 % of variation in global vegetation, as assessed by normalized difference vegetation index data. Further feature inspection showed that the genera Hyphomicrobium, Luedemannella, Reyranella, JGI.0001001.H03, Aeromicrobium, Flavitalea and Ellin6055 most consistently contributed to plant growth prediction. Finally, repeating LASSO regression, an optimized model could predict up to 50 % of variation in global vegetation. In summary, our data suggests a globally conserved set of soil bacterial taxa that correlates with vegetation and might be used to predict plant growth.