Metabolic modeling of microbial communities: Past, present, and future

Main text

Microbial communities are ubiquitous in nature, playing crucial roles in biosphere functions, human health, and industrial processes. Microbial community dynamics and functions result from complex interactions among organisms and between organisms and their environment. Knowledge of such interactions is fundamental to understanding, redirecting, or engineering the function of natural and synthetic communities for various applications. These applications include understanding microbe-host interactions, optimized chemical production, and engineered bioremediation strategies.

Computational models of microbial communities have been critical to the progress in understanding the complex dynamics of microbial ecosystems and rational engineering of stable synthetic microbial communities. Descriptive and predictive computational models of microbial communities can be classified (1) as 1) ordinary-differential-equation-based models of phenotypic traits, 2) models based on sequence-read abundance, 3) agent-based or individual-based models, and 4) genome-scale metabolic models (GEMs). Each of these modeling approaches has unique advantages and limitations. However, GEMs are particularly powerful in describing complex ecosystem interactions. Moreover, they can incorporate knowledge from the other three modeling strategies. Rapid developments in whole-genome sequencing and an abundance of meta-omics data in recent years have improved the prediction quality of GEMs. Recent progress in GEMs has also enabled the investigation of dynamic changes, the impact of perturbations, and the evaluation of species-level interactions.

GEM-based microbial community modeling can be classified based on communities’ dynamics (comprehensively reviewed in (2)): 1) static/unified methods, also known as lumped network methods, typically consider all strains in a common metabolic model with only one copy of the shared reactions and metabolites. They further include strain-specific metabolic content and community biomass target functions. These methods are the simplest approaches, providing a general perspective and allowing for high scalability. For instance, a network-based community model was used to study relationships among the species in a community without considering stoichiometry (3). 2) Static/multi-part methods maintain the individual metabolic matrices that are connected by exchange reactions and introduce a set of metabolites with time-invariant extracellular concentration. For instance, OptCom (4) uses flux balance analysis (FBA) and multi-level and multi-objective optimization at the species and community levels to enable consideration of species-level fitness and community-level objective maximization. 3) Dynamic methods are based on dynamic FBA (dFBA), which allows representations of the community’s temporal changes, metabolite variations, and cell densities over time. While the unified approaches are applicable for multiple strain systems where knowledge is limited, multi-part or dynamic models enable the representation of interspecies interactions and metabolic fluxes. Dynamic models are the only suitable approach to modeling complex situations in microbial communities, such as modeling medium composition, predicting metabolite concentration, and considering time-dependent elements. The dynamic multi-species metabolic modeling (DyMMM; formerly known as DMMM) framework was the first model to use dFBA at the community level and integrate multiple GEMs (5). DyMMM predicts the growth dynamics of multiple species utilizing a single substrate source.

In this issue of the Biophysical Journal, Choudhary and Mahadevan introduce the machine-learning-based DyMMM-LEAPS (DyMMM-locating evenness and stability in large parametric space), an extension of the DyMMM framework (6). As schematically described in Fig. 1, DyMMM-LEAPS explores the large parametric space of genetic circuits in synthetic microbial communities to gain fundamental insights into social interactions and the circuit parameters contributing to evenness and stability. The authors use a machine-learning-based surrogate model trained on adaptive sampling to map the large parametric space of the genetic circuits at a reduced computational cost. Evenness is the extent to which species are equally distributed and is desirable for improving community productivity. Stability indicates a community’s long-term performance and survival, especially when exposed to disturbances. Thus, when designing a synthetic microbial community, it is crucial to ensure both high evenness and stability. The authors simulate cooperation, predation, and competition in coculture and three-strain cultures using syntrophic interactions regulated by quorum-sensing-based genetic circuits. Their results provide a deeper understanding of factors that dictate the evenness and stability of these community interactions.

Figure 1.

A schematic representation of the machine-learning-based dynamic multi-species metabolic modeling-locating evenness and stability in large parametric space (DyMMM-LEAPS) framework along with its inputs and outputs. This methodology enables rapid and computationally efficient exploration of synthetic circuit design space through parametric space reduction without losing critical information.

Developing microbial communities with sustained evenness and stability remains a significant challenge. Computationally efficient models like DyMMM-LEAPS can streamline synthetic circuit design by identifying engineering targets and predicting parameter ranges that yield stable and even microbial consortia, potentially reducing the required experimental permutations. Thus, DyMMM-LEAPS holds promise for advancing the field. It would be important to validate DyMMM-LEAPS by constructing microbial consortia designed by DyMMM-LEAPS and evaluating their evenness and stability against the model’s predictions. Implementing DyMMM-LEAPS recommendations will require leveraging other models, such as RBS calculator v.2.1 (7), for engineering genetic circuits appropriately. Experimental outcomes depend on the accuracy of these models. Moreover, evolutionary robust genetic circuit design principles that allow stability in the absence of selective pressure may also have to be considered (8).

As we advance in systems biology and metabolic modeling, incorporating both biochemical and biophysical temporal changes in environmental conditions (e.g., pH, temperature, stiffness, and porosity) that regulate microbial growth phases would be advantageous. Implementation of stochastic modeling to capture biological noise and other complexities of biological systems would also be valuable. Further development of combined frameworks that incorporate machine and deep learning techniques into GEMs provides an excellent opportunity to elucidate complex biological relationships (9). Lastly, integrating DyMMM-LEAPS with agent-based models that describe the biophysical rules of cell-cell interaction, movement, and emergent behaviors (10) will provide an exciting opportunity to account for spatial heterogeneity in interactions between metabolic networks of individual species. Such integrated computational modeling frameworks can strengthen the link between molecular-scale and population-scale models toward multi-scale modeling from metabolic networks to emergent behaviors.

Declaration of interests

The authors declare no competing interests.

Editor: Guy Genin.