Machine learning reveals biocontrol agents shaping disease outcome in natural Arabidopsis populations

Abstract Plants recruit antagonistic microbes to defend against phytopathogens, offering a route to rational biocontrol beyond empirical screening. Here, using six generations of leaf-microbiome data from natural Arabidopsis populations infected by the oomycete Albugo laibachii , we show that microbial diversity is driven by infection, site, and host genotype, and that infected plants form modular networks with increased inter-kingdom antagonism. We train four machine-learning models to discriminate infected from uninfected plants by microbiota composition and identify microbes enriched in diseased (disease-associated) or healthy (health-associated) plants. Testing the most predictive bacteria, fungi, and cercozoa in planta , we find all confer varying protection against Albugo , with health-associated microbes outperforming disease-associated taxa. The best candidate, a Cystofilobasidium fungus, is validated in a synthetic community, where genomic and community assays indicate biocontrol acts mainly through microbe-microbe interactions rather than plant immune activation. This work shows that pairing microbiome data with machine learning identifies effective biocontrol agents.