Advancing spatial predictions of microbial communities with stacked species distribution models

Abstract

Predicting microbial community composition is essential for belowground biodiversity conservation and management yet remains methodologically challenging. Stacked species distribution models (S-SDM) offer a promising framework by allowing to model large numbers of individual microbial Amplicon Sequence Variants (ASVs) along environmental gradients and assembling them into community-level predictions. However, the consequences of key modelling choices on community-level predictions remain poorly understood, particularly for microbial data. Using bacterial sequence data from 250 topsoil samples collected in the western Swiss Alps through environmentally stratified random sampling, we evaluated how cross-validation and binarisation options influence S-SDM community-level predictions. We assessed prediction accuracy relative to observed communities, differences among predicted communities, and performance along environmental gradients. Stacked ASV model performance in predicting community composition were higher at both extremes of the pH gradient. Predicted communities tended to underestimate compositional dissimilarity but preserved the overall structure of observed beta-diversity. Cross-validation community constraints had negligible effects on predictions. In contrast, binarisation methods strongly influenced outcomes: better community composition predictions were obtained when the sum of predicted probabilities was used to determine threshold value between predicted presence and absence at ASV-level. These results demonstrate that S-SDM can be successfully applied to microbial communities and that the thresholding approach based on the ranking of predicted probabilities yields the most reliable community predictions. Our findings highlight the potential of S-SDMs to support the spatial predictions of microbial biodiversity and their integration into conservation assessments in line with recent IUCN initiatives.