Agricultural practices influence soil microbiome assembly and interactions at different depths identified by machine learning

Soil microbiomes are increasingly recognized for their contribution to crop yields and sustainable agriculture. They enhance nutrient bioavailability for plants, synthesize growth-promoting hormones, and boost plant resilience against stresses, making them valuable for developing new biofertilizer products and informing agricultural management decisions. Agricultural practices, such as tillage, fertilization, and cover cropping, influence soil microbiome structure and function, but their complete impact remains unclear. While environmental selection is a significant factor, stochastic processes might also play a crucial role, especially in later stages of soil ecological succession. A better understanding of how agricultural practices impact soil microbial communities is vital for optimizing soil functionality in agricultural ecosystems. Traditional multivariate methods like PCoA and RDA, while useful, have limitations in analyzing complex soil microbiomes due to their assumptions of linear or unimodal species responses to environmental factors. These methods often fail to capture non-additive effects among environmental factors. Machine learning (ML) offers a powerful alternative, enhancing predictive power and providing insights into species-environment interactions, though often at the cost of reduced model interpretability. This study uses ML, specifically focusing on interpreting complex model predictions using the SHapley Additive exPlanations (SHAP) framework, to investigate microbial assembly, distribution, and networks across depths under different farming practices in a long-term organic field trial. The study aims to identify which abiotic and management factors most strongly influence microbial community assembly in agricultural soils and assess the soil microbiome's ability to signal different agricultural practices. Previous research using linear approaches suggested a strong soil depth signal on microbial community composition but potentially overlooked important patterns.

Existing literature highlights the importance of soil microbiomes in sustainable agriculture and their sensitivity to agricultural practices. Studies have explored how tillage, fertilizers, and cover crops shape microbial communities, with an increasing focus on the balance between deterministic and stochastic processes. While deterministic processes, often linked to environmental variables, influence microbial community assembly, stochastic processes, particularly in later soil succession, may be more impactful than initially thought. Traditional multivariate statistical methods have been widely used to analyze soil microbiome data, but their linear assumptions limit their ability to capture the complexity of these systems. Recent research has demonstrated the potential of machine learning (ML) in analyzing microbiome data, offering improved predictive capabilities and insights into complex interactions. However, ML applications in soil microbiome research remain relatively limited, necessitating further exploration of their utility in understanding microbial assembly and interactions under various agricultural management strategies.

This study used soil samples from the Rodale Institute's Farming Systems Trial (FST), a long-term field trial established in 1981. The FST includes eight different agricultural treatments: three fertility sources (organic legume, organic manure, and synthetic inorganic fertilizer), two tillage types (full and reduced tillage), and the presence or absence of winter cover crops (only for the synthetic fertilizer treatment). Soil cores were collected to a depth of 1 m and divided into four depth sections (0–10 cm, 10–20 cm, 20–30 cm, and 30–60 cm). DNA was extracted, and 16S rRNA and ITS2 regions were sequenced to characterize prokaryotic and fungal communities, respectively. Sequence data were processed using DADA2 to identify amplicon sequence variants (ASVs). Taxonomic annotation was performed using SILVA and UNITE databases. Contaminant ASVs were removed using decontam. A machine learning-based canonical analysis was developed using t-SNE for ordination, providing improved capture of explained variance compared to PCoA, especially for fungal communities. Random forest (RF) models were used to predict ASV abundance based on abiotic and biotic factors (environmental factors, agricultural practices, and microbial interactions). SHAP values were used to interpret RF model predictions. A neutral community model (NCM) was applied to assess the role of stochastic processes in community assembly. Co-occurrence networks were constructed to analyze microbial interactions at different depths. ASVs sensitive to agricultural practices were identified using indicator species analysis and likelihood ratio tests. RF models were used to predict agricultural practices from microbial community data, and a SHAP-based feature selection approach was used to identify primary biomarkers for each agricultural practice. Structural redundancy analysis compared the performance of hub taxa identified by SHAP-based framework and traditional network analysis.

t-SNE ordination revealed that soil depth strongly influenced prokaryotic communities while fertility source was more important for fungal communities. Feature importance analysis using SHAP values showed that fertility source was the most influential factor for fungal communities, while soil depth and organic matter were more important for prokaryotic communities. Tillage significantly affected only prokaryotic communities at shallower depths (0-20 cm). Cover crop effects were less significant and not depth-dependent. The NCM revealed that stochastic processes, particularly dispersal, were more pronounced at shallower depths (0–20 cm), especially in tilled soils. RF models indicated that biotic factors (microbial interactions) were more important than abiotic factors in predicting ASV abundance, particularly at shallower depths. Co-occurrence networks showed greater heterogeneity in microbial interactions across depths, with the highest connectivity in the 30–60 cm layer. Fertility source-sensitive ASVs were concentrated in the top 30 cm, while tillage-sensitive ASVs were primarily in the top 20 cm. Modules dominated by fungi sensitive to fertility source were identified at depths up to 30 cm. A large proportion of ASVs within fertility modules were predictable from other ASVs in the module, emphasizing fungi's crucial role. SHAP analysis identified fungi as major hub taxa in fertility modules. The model demonstrated high accuracy in differentiating fertility sources, tillage, and cover crops based on microbial community composition. Key fungal biomarkers for fertility sources included ASVs from families Nectriaceae, Chaetomiaceae, and Pyronemataceae, predominantly enriched in organic fertility sources. Prokaryotic biomarkers showed more even distribution across fertility sources. Tillage-enriched taxa included fungal families Lasiosphaeriaceae and Didymellaceae and prokaryotic families Nitrosomonadaceae and Gemmatimonadaceae. Biomarkers for cover crops were less pronounced.

This study provides insights into the drivers of soil microbiome assembly and interactions under various agricultural practices. Machine learning methods, particularly t-SNE and SHAP, revealed patterns not easily detectable with traditional linear approaches. The strong influence of fertility source on fungal communities, particularly at shallower depths, highlights the importance of organic matter inputs in shaping fungal communities. The higher importance of soil depth for prokaryotic communities underscores the differing responses of prokaryotes and fungi to environmental gradients. The influence of tillage on shallow soil microbiomes by increasing stochasticity, especially in prokaryotic communities, points to the disruptive effects of tillage on microbial interactions. The minimal and inconsistent influence of cover crops on microbial communities, irrespective of depth, suggests the need for further investigation into optimal cover cropping strategies to maximize microbiome benefits. The findings support the use of ML methods to disentangle complex interactions within soil microbial communities and provide new avenues for managing soil health through sustainable agricultural practices.

This research demonstrates the use of machine learning to gain a deeper understanding of soil microbiome assembly and interactions under different agricultural practices. The study highlights the strong influence of fertility source, particularly on fungal communities, the depth-dependent effect of tillage on microbial dispersal and stochastic processes, and the relatively minor and inconsistent influence of cover cropping. Future research could explore finer-scale spatial heterogeneity of soil properties and microbial communities to improve model predictions and investigate the functional roles of identified biomarkers in contributing to soil health and crop yields. Further research could also examine the long-term effects of different agricultural practices on microbiome stability and resilience.

One limitation is the potential underestimation of the impact of agricultural practices due to the correlation between practices and soil properties. The analysis focused primarily on bulk soil properties and might not fully capture the intricate spatial heterogeneity of microbial communities at finer scales. The neutral community model, while valuable for assessing stochastic processes, does not fully account for the complex biotic and abiotic interactions within soil microbial communities. The study’s focus on a single long-term field trial might limit the generalizability of its findings to other geographic locations or soil types. Further investigation is required to fully understand the functional roles of identified biomarkers in relation to agricultural practices and soil health.