Abstract
Soil-borne plant diseases are increasingly causing devastating losses in agricultural production. The development of a more refined model for disease prediction can aid in reducing crop losses through the use of preventative control measures or soil fallowing for a planting season. The emergence of high-throughput DNA sequencing technology has provided unprecedented insight into the microbial composition of diseased versus healthy soils. However, a single independent case study rarely yields a general conclusion predictive of the disease in a particular soil. Here, we attempt to account for the differences among various studies and plant varieties using a machine-learning approach based on 24 independent bacterial data sets comprising 758 samples and 22 independent fungal data sets comprising 279 samples of healthy or Fusarium wilt-diseased soils from eight different countries. We found that soil bacterial and fungal communities were both clearly separated between diseased and healthy soil samples that originated from six crops across nine countries or regions. Alpha diversity was consistently greater in the fungal community of healthy soils. While diseased soil microbiomes harbored higher abundances of Xanthomonadaceae, Bacillaceae, Gibberella, and Fusarium oxysporum, the healthy soil microbiome contained more Streptomyces Mirabilis, Bradyrhizobiaceae, Comamonadaceae, Mortierella, and nonpathogenic fungi of Fusarium. Furthermore, a random forest method identified 45 bacterial OTUs and 40 fungal OTUs that categorized the health status of the soil with an accuracy >80%. We conclude that these models can be applied to predict the potential for occurrence of F. oxysporum wilt by revealing key biological indicators and features common to the wilt-diseased soil microbiome.