Soils are a heterogeneous environment, with complex dynamic interactions between spatio-temporally varying biotic and abiotic parameters that determine the organization of soil microbial communities. Measuring microbial community structure can often result in many and complex measures of composition, abundance and activities. To understand the relationship between microbial community structure and function often requires an appreciation of the inherent multivariate structure on both sides of the relationship, elucidated by a variety of statistical and computational tools.

We can explore complex data with multivariate methods such as cluster analysis and stepwise discriminant analysis. Discrimination by genomic community structure measures was considerably more successful than general measures of abundance and activity applied to soil classes defined by chemical variables. One can view the soil environment characteristics as the driving factors and the soil microbial community structure as responses in a factor-response model of the soil system response to management changes. We successfully constructed a series of models of association using confirmatory factor analysis or structural equation modeling.

To discriminate between soil environments, we have designed computational tools based on Support vector machines (SVM) and K-nearest neighbors (KNN), which are two machine learning tools, to perform supervised classification of soil samples based on chemical composition and constituent microbial profiles. They were tested on soil samples from Idaho, Chesapeake Bay, and Miami-Dade County. Unsupervised classification tools based on clustering and self-organizing maps to classify unknown soil samples into similar groups have also been designed.

Microbial communities in soil vary in composition, abundance and activities. Key factors influencing the communities are different at different spatial scales, often resulting in different conclusions. Understanding the scaling effect on many of the conceptual and computational models linking structure to function will be a key challenge for the future.