Abstract
Soil microbiomes are critical to soil health and sustainable agriculture, yet their assembly and distribution across different soil depths under varying management practices remain poorly understood. This study applied machine learning-based methods to identify prokaryotic and fungal community patterns across soil depths (0–60 cm) in fields managed with different tillage, fertility sources, and cover crops. Results indicate that fertility source is the primary driver of microbial assembly, particularly for fungi, though its impact diminishes with depth. Tillage was found to predominantly affect the top 20 cm, enhancing microorganism dispersal and amplifying stochastic processes, particularly in prokaryotic communities. Notably, machine learning models demonstrated that biotic interactions are significantly stronger predictors of microbial abundance than abiotic environmental factors. These findings highlight the multifaceted impact of agricultural practices on microbial assembly and the power of machine learning to decode complex ecological interactions.
Introduction
Sustainable agriculture relies heavily on the functional stability of soil microbiomes. However, soil is often treated as a two-dimensional surface in research, overlooking the fact that microbial communities change drastically with depth. Management practices like tillage, fertilization, and cover cropping alter these communities, but traditional statistical methods often struggle with the high-dimensional, sparse, and non-linear nature of microbiome data. This study addresses these gaps by employing robust predictive frameworks to understand how management decisions influence microbial assembly and interaction networks from the surface down to 60 cm.
Methods
Researchers utilized data from the Rodale Institute’s Farming Systems Trial (FST), comparing organic (legume and manure) and conventional (synthetic fertilizer) systems. Soil cores were analyzed at four depth increments: 0–10, 10–20, 20–30, and 30–60 cm. To move beyond linear statistics, the study employed:
- Manifold Learning (t-SNE): To capture non-linear relationships in microbial ordination.
- Random Forest (RF) & SHAP: To predict microbial abundance and identify key biomarkers using the SHapley Additive exPlanations framework.
- Neutral Community Model (NCM): To distinguish between deterministic processes (environmental selection) and stochastic processes (random dispersal and birth/death events).
- Co-occurrence Networks: To identify "hub taxa" that exert disproportionate influence on community structure.
Key Insights and Results
Tillage as a Randomizer
Tillage significantly altered the assembly processes in the top 20 cm of soil. By physically mixing the soil, tillage enhanced microorganism dispersal but also amplified stochasticity. The Neutral Community Model (NCM) revealed that random processes played a larger role in these shallower depths, potentially jeopardizing the stable biotic interactions found in undisturbed soil.
Fertility Source and Depth Sensitivity
The type of fertility source (organic vs. synthetic) was the most pronounced driver of microbial assembly, especially for fungal communities. While the influence of tillage and cover crops was largely confined to the 0–20 cm range, the signal from fertility sources remained detectable down to 60 cm. Organic fertility sources fostered sensitive modules dominated by fungi, suggesting a more integrated microbial network compared to synthetic treatments.
The Dominance of Biotic Interactions
A pivotal finding of the machine learning analysis was that "biotic factors"—the presence and abundance of other microbes—were much stronger predictors of a specific microbe's abundance than "abiotic factors" like soil pH, carbon, or nitrogen. This suggests that microbial "social networks" are the primary stabilizers of community structure.
Conclusions
Agricultural management practices impact soil microbiomes through a complex interplay of assembly processes, interactions, and the recruitment of specific taxa. Tillage acts as a randomizing force in topsoil, while fertility sources shape long-term community architecture across the soil profile. By leveraging machine learning, this research demonstrates that microbial interactions are more predictive of community composition than environmental variables alone, offering a new lens through which to optimize soil health for sustainable agriculture.
Summary updated