Explainable machine learning to map context-dependent biotic controls on soil organic carbon persistence across global pedoclimatic domains

Soil organic carbon (SOC) assessments often prioritize climate and vegetation, while the biological regulation of organic matter formation and persistence remains underrepresented. Our aim was to develop an explainable machine learning framework to model SOC stocks and identify hotspots where roots and fungi influence SOC persistence mechanisms. We combined climatic variables, physicochemical and mineralogical soil indicators, and biological predictors related to roots and fungi. A Quantile Random Forest model was trained, and predictions were interpreted with Shapley values and local dominance spatial maps. Temperature and precipitation showed patterns: SOC peaked in cold conditions (0-10 °C) and increased with annual precipitation up to ~4000 mm yr^-1. However, soil biota predictors introduced contextual dependence in the inferred persistence mechanisms. Root-fungus symbiosis emerged as a key factor: high mycorrhizal colonization aligned with higher SOC in humid, clay-rich environments, while similar biological conditions did not compensate for SOC vulnerability in dry, sandy environments. This contrast across pedoclimatic domains indicates that biotic influences depend on the physical template governing protection, aeration, and water limitation. Dominance maps revealed nitrogen contrasts: in arid and semi-arid regions, N predominated in explaining negative SOC anomalies, while in cold, humid environments, N predominated in explaining positive anomalies. This does not imply a simple causal relationship; instead, the marginal contribution of N is spatially rearranged in covariation with climate, texture, and other soil attributes, partly reflecting the co-occurrence between organic N and SOC. Overall, explainable modeling helps identify where biodiversity most strongly influences SOC persistence and supports decision-making