Modeling interaction networks between host, diet, and bacteria predicts obesogenesis in a mouse model

Abstract

Host-microbiome interactions are known to have substantial effects on human health, but the diversity of the human microbiome makes it difficult to definitively attribute specific microbiome features to a host phenotype. One approach to overcoming this challenge is to use animal models of host-microbiome interaction, but it must be determined that relevant aspects of host-microbiome interactions are reflected in the animal model. One such experimental validation is an experiment by Ridura et al. In that experiment, transplanting a microbiome from a human into a mouse also conferred the human donor's obesity phenotype. We have aggregated a collection of previously published host-microbiome mouse-model experiments and combined it with thousands of sequenced and annotated bacterial genomes and metametabolomic pathways. Three computational models were generated, each model reflecting an aspect of host-microbiome interactions: 1) Predict the change in microbiome community structure in response to host diet using a community interaction network, 2) Predict metagenomic data from microbiome community structure, and 3) Predict host obesogenesis from modeled microbiome metagenomic data. These computationally validated models were combined into an integrated model of host-microbiome-diet interactions and used to replicate the Ridura experiment in silico. The results of the computational models indicate that network-based models are significantly more predictive than similar but non-network-based models. Network-based models also provide additional insight into the molecular mechanisms of host-microbiome interaction by highlighting metabolites and metabolic pathways proposed to be associated with microbiome-based obesogenesis. While the models generated in this study are likely too specific to the animal models and experimental conditions used to train our models to be of general utility in a broader understanding of obesogenesis, the approach detailed here is expected to be a powerful tool of investigating multiple types of host-microbiome interactions.

Keywords: computational modeling; metabolome; microbiome; mouse model; network biology; obesity.

FIGURE 1

A community interaction network for mouse microbiome communities. This figure shows the result from the MAP-model. Diamonds are diet parameters, amino acids are blue, carbohydrates are yellow, fats are green, minerals are purple, and vitamins are orange. Circles are bacterial taxa and the size of bacterial node is proportionate to their average relative abundance across all analyzed microbiomes. Solid lines indicate interaction between taxa at final time point, dashed line indicate interactions between taxa at time initial and final time point.

FIGURE 2

Predictions for change in microbiome in response to host diet. Y-axis is average PCC between predicted and observed mouse gut microbiome community structures for randomly generated subsets of training and validation data. On x-axis is the two modeling approaches considered: “Network-based” and “Non-network Based” models. Average results for Training and Test data subsets are shown. Error bars are +/− one standard deviation.

FIGURE 3

Predict microbiome enzyme function profiles from community structures. On x-axis is the two modeling approaches considered: “Initial Prediction” and “Optimized Prediction”Y-axis is the average PCC between predicted and observed mouse gut EFPs. Error bars are +/− one standard deviation.

FIGURE 4

Predictions for host diet and obesity. In the top graph, results for prediction of host diet from microbiome data are shown. In the bottom graph, results for prediction of host obesity state from microbiome data are shown. Y-axis is MCC score for binary classification quality. The nature of the data used, “Microbial Community Structure” or “Microbial Community Metabolome”, to train the models are listed on the X-axis. Results are presented for both training and validation data subsets.

FIGURE 5

Results of in silico microbiome transplant experiment. The in silico results of the microbiome experiment, in which mice with a “Lean” or “Obese” starting microbiome are fed a Low Fat (LF) or High Fat (HF) diet. Y-axis is the predicted obesogenesis of host-microbiome-diet interactions with larger values indicated increased obesogenesis.