Understanding the effectiveness and potential mechanism of action of agricultural biological products under different soil profiles and crops will allow more precise product recommendations based on local conditions and will ultimately result in increased crop yield. This study aimed to use bulk soil and rhizosphere microbial composition and structure to evaluate the potential effect of a Bacillus amyloliquefaciens inoculant (strain QST713) on potatoes and to explore its relationship with crop yield. We implemented next-generation sequencing (NGS) and bioinformatics approaches to assess the bacterial and fungal biodiversity in 185 soil samples, distributed over four different time points-from planting to harvest-from three different geographical locations in the United States. In addition to location and sampling time (which includes the difference between bulk soil and rhizosphere) as the main variables defining the microbiome composition, the microbial inoculant applied as a treatment also had a small but significant effect in fungal communities and a marginally significant effect in bacterial communities. However, treatment preserved the native communities without causing a detectable long-lasting effect on the alpha- and beta-diversity patterns after harvest. Using information about the application of the microbial inoculant and considering microbiome composition and structure data, we were able to train a Random Forest model to estimate if a bulk soil or rhizosphere sample came from a low- or high-yield block with relatively high accuracy (84.6%), concluding that the structure of fungal communities gives us more information as an estimator of potato yield than the structure of bacterial communities. IMPORTANCE Our results reinforce the notion that each cultivar on each location recruits a unique microbial community and that these communities are modulated by the vegetative growth stage of the plant. Moreover, inoculation of a Bacillus amyloliquefaciens strain QST713-based product on potatoes also changed the abundance of specific taxonomic groups and the structure of local networks in those locations where the product caused an increase in the yield. The data obtained, from in-field assays, allowed training a predictive model to estimate the yield of a certain block, identifying microbiome variables-especially those related to microbial community structure-even with a higher predictive power than the geographical location of the block (that is, the principal determinant of microbial beta-diversity). The methods described here can be replicated to fit new models in any other crop and to evaluate the effect of any agricultural input in the composition and structure of the soil microbiome.
Keywords: agricultural biological; machine learning; soil microbiome; yield prediction.