Machine learning-based mapping of Acidobacteriota and Planctomycetota using 16 S rRNA gene metabarcoding data across...

The soil microbiome plays a crucial role in maintaining healthy ecosystems and supporting sustainable agriculture. Studying its biogeographical structure and distribution is essential for understanding the rates and mechanisms of microbially mediated soil ecosystem services. This study aimed to investigate the spatial distribution patterns of Acidobacteriota and Planctomycetota across soils in Russia, summarizing data from 16S rRNA gene metabarcoding of topsoils. A machine learning approach (Random Forest) was employed to generate digital distribution maps using climatic, topographic, vegetation, geological, and soil variables. Model interpration was performed using variable importance assessment and Shapley values. According to the error metrics, the Acidobacteriota model achieved a root mean squared error (RMSE) of 6.67% and an R² of 0.41, while the Planctomycetota model achieved an RMSE of 2.04% and an R² of 0.46. Both phyla exhibited similar spatial distribution patterns, with relative abundance decreasing from North to South. For Acidobacteriota, vegetation cover, surface temperature, and soil pH were significant predictors, whereas the relative abundance of Planctomycetota was mainly influenced by climatic variables. Specifically, Acidobacteriota were more abundant in areas with dense vegetation, stable surface temperatures, and acidic soils. In contrast, Planctomycetota showed reduced abundance in regions with higher levels of precipitable water vapor. These results highlight the potential of machine learning techniques to visualize predictive biogeographic patterns in soil microbial taxa abundance at the phylum level. Despite limitations related to the heterogeneous nature of source data, focusing on higher taxonomic ranks less sensitive to methodological variation enabled to identify preliminary large-scale distribution trends of microbial phyla in soils.