From model to field: predictive design of microbial communities for resilient crop performance

Abstract

Synthetic microbial communities (SynComs) represent a promising approach to enhance crop growth and stress resilience through microbiome engineering. However, the systematic design and field validation of SynComs remain limited. Here, we present a predictive framework for SynCom optimization, integrating plant phenotyping, microbial genomics, and machine learning. Using tomato as a model, we tested over 800 SynCom-temperature combinations consisting of root endophytic bacteria and rhizosphere metabolites. An Elastic Net regression model trained on plant biomass data accurately predicted the performance of unseen SynComs, with prediction accuracy plateauing at [~]5% (301/6144) of all possible SynCom-temperature combinations. Incorporating genomic features significantly improved model performance, whereas microbiome compositional data alone were not informative. We applied the model to design novel SynComs, which were tested in both laboratory and field conditions using a commercial tomato cultivar. The model-informed SynCom enhanced plant growth in field trials and improved heat stress tolerance under controlled laboratory conditions. Multi-omics analyses and feature importance metrics identified specific microbial taxa, including Sphingobium sp., whose enrichment was linked to host plant metabolite (e.g., tomatine) and stress-responsive gene expression. Our results demonstrate a scalable strategy for the predictive design of beneficial microbiomes to improve resilient crop performance under real-world conditions.