An Interpretable Machine Learning Approach to Ecologically Characterize Soil Carbon and Structure From Multi-Kingdom...

Soil physical structure is a critical determinant of agricultural landscape resilience, yet standard pedotransfer functions estimate soil hydraulic and structural properties using static abiotic variables, often overlooking the biological mechanisms that actively organize soil structure. This study evaluates the predictive power of multi-kingdom microbiome data (prokaryotes, fungi and microeukaryotes) for three key soil functions: soil organic carbon (SOC) stock, mean weight diameter (MWD) and macroporosity. Using a dataset of 2251 agricultural soil samples from Quebec, Canada, we benchmarked four machine learning algorithms (HGBR, RFR, XGBoost, SVR) and four data aggregation strategies. The integration of microbiome data with texture and climate variables achieved high peak predictive accuracy ( $R^2$ range: 0.70-0.82). Methodologically, high-resolution compositional approaches (ASV-level centered log-ratio) and kingdom-balanced absolute abundances consistently outperformed taxonomic or functional aggregations. The loss of predictive power at the family level indicates that traits governing soil physical modification are phylogenetically shallow and strain-specific. Interpretability analysis using Shapley Additive Explanations (SHAP) revealed a clear functional hierarchy in soil assembly. Specific prokaryotic and fungal features drove biochemical stabilization and physical scaffolding via the microbial carbon pump and structural enmeshment dynamics. In contrast, the architectural openness of macroporosity was fundamentally constrained by abiotic physical limits (e.g., texture). Within this physical framework, specific microbial taxa, including anaerobic bacteria and microeukaryotic amoebae, functioned not as active engineers, but as high-sensitivity bio-indicators of the resulting aeration and hydrological connectivity. These results define soil physical organization as a biologically mediated hierarchy rather than a passive geological byproduct. Consequently, we propose shifting from static pedotransfer functions to a dynamic biotransfer framework that leverages multi-kingdom omic signatures to monitor soil physical resilience and crop adaptation potential.