Predicting soil microbiome structure and function using genome-scale metabolic modeling

Background/Question/Methods

Nitrogen (N) limits important ecosystem processes such as productivity and decomposition, and biogeochemical models can be used to track the fluxes of N between pools. However, these models often underperform under changing conditions, especially for microbial processes (e.g. nitrification, ammonification). An approach called genome-scale metabolic modeling, initially developed for controlled lab conditions, offers the potential to simulate bacterial activity under dynamic environmental conditions. We compare this method for predicting soil N fluxes against simpler approaches for integrating microbes into models, such as using marker gene counts. We evaluate its ability to capture emergent components of soil biotic structure, such as soil depth gradients of the Nitrospira genus and abundances of nitrifiers across temperate forest soils. Simulations used dynamic flux-balance analysis with the spatially-explicit population-based COMETS modeling framework. To approximate a soil profile, simulations were initialized with vertical diffusion of oxygen and carbon dioxide and randomly-located populations of Nitrospira species. Community simulations of 16 soil taxa (to predict nitrifier abundances and N fluxes) were initialized with soil NO3, NH4, and moisture from Harvard Forest samples collected by the National Ecological Observatory Network, and then evaluated against microbial and N flux data from those samples.

Results/Conclusions

We show that simulations using genome-scale metabolic network models, representing a subset of the complex soil microbiome, can capture major trends in soil N cycling. Observed N flux rates were more strongly correlated with simulation outputs than with genomic data on marker genes, especially when paired with soil properties; this improvement over non-microbial models was stronger for nitrification than for ammonification, reducing root mean squared error (RMSE) by 27% and 4%, respectively. Ammonification was better explained overall than nitrification (ammonification R2= 0.37, nitrification R2= 0.32). Simulations with oxygen and CO2gradients were also able to recreate depth gradients of Nitrospira species, with the majority of microbial biomass concentrated in the deepest 30% of the soil profile. Simulations across NEON sites also successfully predicted abundances of nitrifying bacteria (R2= 0.47, p< 0.001). We conclude that genome-scale models have the capacity to predict microbial community dynamics and function in complex soil environments. These results illustrate a promising avenue for incorporating computational genomics into microbial ecology and soil biogeochemistry.