Machine Learning Reveals Biocontrol Agents Shaping Disease Protection in Natural Arabidopsis Populations and Synthetic Communities

Plants recruit beneficial microbes to defend against phytopathogens, and some are developed as biocontrol agents. However, identifying biocontrol agents depends on empirical screenings to assess microbial activity against pathogens of interest. In this study, we explored the association between infection status and phyllosphere microbiome dynamics over six generations from natural Arabidopsis populations and compared plants infected by the oomycete pathogen Albugo laibachii with uninfected controls. We found that infected plants exhibited reduced microbial diversity, a pattern driven primarily by site-dependent environmental factors rather than host genotypes. Microbial interaction networks of infected plants were less connected and highly modular, indicating a disruption in microbial community resilience. Using machine learning, we predicted the disease-associated and health-associated microbial signatures with 86–91% accuracy and identified key taxa correlated with disease outcomes. We selected bacteria, fungi, and cercozoa from these key taxa to test their biocontrol activeity against A. laibachii infection. We found all selected microbes exhibit various levels of plant protection, and disease-associated microbes offered less protection than health-associated microbes. We further validated the biocontrol capacity of the best candidate agent, Cystofilobasidium, by testing it within a synthetic community derived from the Arabidopsis core microbiome. This study reveals how microbial dynamics and functional connectivity support plant protection, providing a roadmap for using machine learning to devise robust biocontrol strategies that enhance crop resilience.