Dynamic simulations of microbial communities under perturbations: opportunities for microbiome engineering

Background : There are few large longitudinal microbiome studies, and fewer that include controlled, well-annotated perturbations between sampling-points. Thus, there are few opportunities to employ data-driven computational analyses of perturbed microbial communities over time. Results : Our novel computational system simulates the dynamics of microbial communities under perturbations using genome-scale metabolic models (GEMs). Perturbations include modifications to a) the nutrients available in the medium, allowing modelling of prebiotics; and/or b) the microorganisms present in the community to model, for example, probiotics or pathogen infection. These simulations generate the quantity and types of information required by MDPbiome, an AI system which builds predictive models suggesting the perturbation(s) required to engineer microbial communities to a desired state. We call this novel combination of technologies "MDPbiomeGEM"'. We demonstrate, in a Crohn’s disease microbiome, that MDPbiomeGEM correctly models the influence of both prebiotic fiber and a probiotic, resulting in a recommendation to consume inulin to recover from dysbiosis, consistent with prior biomedical knowledge. When used to model the soil microbiome's ability to degrade the herbicide Atrazine, differing recommendations arise depending on the highly variable state of the initial soil microbial composition, highlighting the relevance of both phosphate and microbes (i.e. Halobacillus sp. and H.stevensii ) in a directed microbiome engineering strategy, consistent with previously published observations. Conclusions : MDPbiomeGEM generates large volumes of longitudinal data of complex microbial communities experiencing perturbations. Machine learning on these data reveal patterns consistent with existing biological knowledge, supporting the validity of the approach. MDPbiomeGEM could save research resources by optimizing sample collection in metagenomics studies through identification of "informative" scenarios/time-points, or by predicting optimal in-vitro culture formulations for generating performant synthetic microbial communities. Finally, MDPbiomeGEM outputs include detailed information about the metabolic state of the community, which can be used to further interpret the impact of perturbations, and potentially could be used to predict novel metabolic biomarkers of a microbiome's


Abstract
Background : There are few large longitudinal microbiome studies, and fewer that include controlled, well-annotated perturbations between sampling-points. Thus, there are few opportunities to employ data-driven computational analyses of perturbed microbial communities over time.
Results : Our novel computational system simulates the dynamics of microbial communities under perturbations using genome-scale metabolic models (GEMs). Perturbations include modifications to a) the nutrients available in the medium, allowing modelling of prebiotics; and/or b) the microorganisms present in the community to model, for example, probiotics or pathogen infection. These simulations generate the quantity and types of information required by MDPbiome, an AI system which builds predictive models suggesting the perturbation(s) required to engineer microbial communities to a desired state. We call this novel combination of technologies "MDPbiomeGEM"'. We demonstrate, in a Crohn's disease microbiome, that MDPbiomeGEM correctly models the influence of both prebiotic fiber and a probiotic, resulting in a recommendation to consume inulin to recover from dysbiosis, consistent with prior biomedical knowledge. When used to model the soil microbiome's ability to degrade the herbicide Atrazine, differing recommendations arise depending on the highly variable state of the initial soil microbial composition, highlighting the relevance of both phosphate and microbes (i.e. Halobacillus sp. and H.stevensii ) in a directed microbiome engineering strategy, consistent with previously published observations.
Conclusions : MDPbiomeGEM generates large volumes of longitudinal data of complex microbial communities experiencing perturbations. Machine learning on these data reveal patterns consistent with existing biological knowledge, supporting the validity of the approach. MDPbiomeGEM could save research resources by optimizing sample collection in metagenomics studies through identification of "informative" scenarios/time-points, or by predicting optimal in-vitro culture formulations for generating performant synthetic microbial communities. Finally, MDPbiomeGEM outputs include detailed information about the metabolic state of the community, which can be used to further interpret the impact of perturbations, and potentially could be used to predict novel metabolic biomarkers of a microbiome's state.
3 Full Text Due to technical limitations, full-text HTML conversion of this manuscript could not be completed.
However, the manuscript can be downloaded and accessed as a PDF.

Tables
Due to technical limitations, the tables are only available as a download in the supplemental files section. Figure 1 MDPbiomeGEM general schema. MMODES uses GEM models (left) to generate microbial community synthetic, dynamic data (center, for example Fig. 2a), that are in turn consumed by MDPbiome (right), and used to generate a directed intervention recommendation. with recommended interventions to avoid, or recover from, Crohn's disease dysbiosis 5 highlighted in red. `Pro-high/low' means F.prausnitzii as a probiotic in high/low concentration.

Figure 3
Stability and generality of the MDPbiome recommendations to recover from Crohn's disease.
a Stability assessment: rows represent the three microbiome states. Within each row, the color fragments represent each of the defined perturbations. The darker color fragment represents the optimal recommendation for each state (one per row). b Generality assessment: frequency of transitions when following (row F), or not following (row nF), the optimal recommendation. Better (green), equal (blue) and worse (red) state-transition are defined based on the utility function for sorting states -in this case, higher butyrate concentration.
6 Figure 4 Metabolites change in their concentration ratio after perturbation. Average of samples from all microbiome states after a particular perturbation was applied. Salmon indicates no change in concentration; yellow and light orange indicate a decrease; blue colors indicate an increase. 2.5 was established as maximum possible relative concentration ratio  Atrazine degradation results. a Atrazine degradation ratio (black dots) is computed as the percentage of atrazine that is degradaded between two sampling points. It is computed by sample (several samples per time series), and may be aggregated by state (mean of samples within that state) or by perturbation (mean of ratio of samples after a specific perturbation is applied). b Relative probability to move between the states. c Stability assessment (described in Fig.3a). Hb=Halobacillus sp., Hm=H. stevensii, Hb-Hm=both.

Supplementary Files
This is a list of supplementary files associated with this preprint. Click to download.