Modeling recovery of Crohn’s disease, by simulating
microbial community dynamics under perturbations
Jorge Carrasco Muriel, Beatriz Garc´ıa-Jim´enez and Mark D. Wilkinson
Biological Informatics Group, Center for Plant Biotechnology and Genomics, UPM - INIA,
Universidad Polit´ecnica de Madrid, 28223 (Madrid), Spain
email@example.com, ORCiD : 0000-0002-8129-6506
There are few large longitudinal microbiome studies, and fewer that include planned, annotated perturbations between
sampling-points. Thus, there are few opportunities to employ data-driven computational analyses of perturbed microbial
communities over time.
Our novel computational system simulates the dynamics of microbial communities under perturbations, using genome-
scale metabolic models (GEM). Perturbations include modiﬁcations to a) the nutrients available in the medium, allowing
modelling of prebiotics; or b) the microorganisms present in the community, to model probiotics or pathogen infection.
These simulations generate the quantity and types of information used as input to the MDPbiome system, which builds
predictive models suggesting the perturbation(s) required to engineer microbial communities to a desired state.
We demonstrate that this novel combination, called MDPbiomeGEM, is able to model the inﬂuence of prebiotic ﬁber
and probiotic in the case of a Crohn’s disease microbiome. The output’s recommended perturbation to recover from
dysbiosis is to consume inulin, which promotes butyrate production to reach homeostasis, consistent with prior biomed-
ical knowledge. Our system could also contribute to design (perturbed) microbial community dynamics experiments,
potentially saving resources both in natural microbiome scenarios by optimizing sequencing sampling, or to optimize
in-vitro culture formulations for generating performant synthetic microbial communities.
Whilst the static microbial community composition in a multitude of niches have been extensively analyzed, studies of the
dynamics of a microbiome are scarce, or focused on very few sampling points; for example, where the eﬀect of a treatment
versus a control are compared. Thus, few long-term metagenomics studies are available to illustrate the variability of
microbial communities over time, and even fewer with suﬃcient meta-data regarding events that could inﬂuence microbial
If such large longitudinal datasets were available, computational analysis could help to understand microbial community
dynamics and, indeed, provide insights on how to modify it in desirable ways. If there were suﬃcient samples, data-
driven computational systems, such as MDPbiome (5), could provide suggestions to engineer the microbiome, guiding the
microbial community to a healthier or more performant state. MDPbiome is an Artiﬁcial Intelligence system, built using
a Markov Decision Process, to provide in-silico recommendations (e.g., about diet, pre/pro-biotics, drugs, fertilizers, etc.)
to guide the subject’s microbiome through a path towards health or high performance, relying on microbial community
changes in response to perturbations.
A Genome Scale Model (GEM) is a metabolic reconstruction of a speciﬁc organism, that can be transformed to a
mathematical model, with Flux Balance Analysis (FBA) being the most widely-applied method to solve such models,
using Linear programming to optimize cell growth. GEM was originally designed to simulate the metabolism of one
isolated strain. In the last years, however, some novel approaches are emerging that combines assorted GEMs at a
community level. However, there are several challenges to address such as: a) the lack of standardization (diﬀerent models
from diﬀerent authors using diﬀerent nomenclatures, b) deﬁning medium composition, and c) determining exchange rates
to constrain the ﬂuxes in a community context.
2 Our proposal
We propose a combination of MDPbiome and GEM (MDPbiomeGEM), to simulate the dynamics of a (simpliﬁed)
microbial community under perturbations using GEMs, in order to generate suﬃcient longitudinal data to enable an
MDPbiome-based analysis. MDPbiomeGEM identiﬁes diﬀerent microbial community states, and the transitions between
them, guided by perturbations. Finally, it computes a policy recommendation to move the microbiome from a diseased
or low-performance state to a healthy or high-performance one.
There are several categories of studies that use GEM for microbial community modeling. Brieﬂy, these can be simpliﬁed
into static and dynamic approaches. Static methods make up the majority of studies (e.g. MO-FBA, (d)OptCom, cFBA,
CASINO, Mminte, CarveMe, Microbiome modeling toolbox, etc.), with dynamic approaches being rarer and not widely
used (e.g. COMETS, BacArena, Daphne, etc.). Only dynamic methods allow the simulation of microbial community
evolution over time, including explicit metabolite exchange and variability of the metabolite concentration in the medium,
as are the requirements for our simulation system. There are few studies combining metabolic models and microbiome
analyses (Thiele et al., SteadyCom, (6)), although these pursue primarily a static approach, in contrast the dynamic
approach suggested here.
Following a dynamic approach, we develop a novel dynamic community GEM approach: MMODES (Metabolic Models
based Ordinary Diﬀerential Equations Simulation). We select the most complex (and little-studied) combination of
characteristics for a GEM in order to be able to simulate a complex microbial community, satisfying physiological knowledge
as much as possible. The simulated community can be characterized as follows: a) multi-strain; b) dynamic (time
dependent); c) hybrid, updating biomass and metabolite concentration over time; and d) experiences perturbations.
MMODES is based on Daphne (7). Our main distinguishing contributions are: 1) the inclusion of perturbations, 2)
concentration updates of speciﬁc metabolites, rather than ad hoc ones, 3) a simpliﬁed human interface. MMODES
manages the individuals’ GEMs with Constraints Based Reconstruction and Analysis (COBRA), and solves them with
(p)FBA. The dynamics is implemented in a similar way to that of dynamicFBA, by alternating a phase of solving individual
models with a phase of updating of the metabolite concentrations in the medium and the strain biomasses.
To validate our system, we selected the microbial community and GEMs described by (1), composed of representatives of
two of the three dominant phyla in the human gut (Actinobacteria and Firmicutes): Biﬁdobacterium adolescentis L2-32
(iBif452) and Faecalibacterium prausnitzii A2-165 (iFap484). Both strains have a cross-feeding relation through acetate.
Our goal is to model the dynamics of the human gut microbiome associated with Crohn’s disease, such as a depletion of
F.prausnitzii as is symptomatic of this disease (4).
2 4 6
Inulin:0.19; FOS:0.78; Starch:1
Inulin:0.81 FOS:0.22 Pro-high:0.25 Pro-low:0.11 Starch:0.37
. Pro-high:0.64; Pro-low:0.78
- Inulin:0.39; FOS:0.95; Starch:0.63
Inulin:0.61 FOS:0.05 Pro-high:0.36 Pro-low:0.22 Starch:0.66
Inulin:1; FOS:1; Starch:0.34
A B C
Figure 1: MDPbiomeGEM modeling Crohn’s disease recovery A) Example of one microbial community timeseries
with perturbations, generated by MMODES. B) Sequence of microbiome states followed by multiple timeseries (one
per row). C) MDPbiome state-transition diagram, with recommended perturbations to avoid Crohn’s diseases,
highlighted in red. ‘Pro-high/low’ means F.prausnitzii as a probiotic in high/low concentration.
With MMODES, we simulate multiple timeseries (e.g. Fig. 1A), corresponding to diﬀerent initial relative abundances
of the two microbes in the community, covering the full range of possibilities (from healthy to dysbiosis; i.e. higher to
lower proportions of F.prausnitzii), and with random perturbations inﬂuencing the dynamics of the community. The
perturbations considered are diﬀerent ﬁber sources (considered prebiotics) that feed one or both strains, speciﬁcally: a)
inulin (only F.prausnitzii ferments it), b) starch (only B.adolescentis ferments it) and c) FOS (both strains consume
it). The GEMs are adjusted to satisfy these fermentation capabilities checked experimentally (2). As an additional
perturbation, we also model the intake of F.prausnitzii as a probiotic, simulating a low/high increase in the biomass of
that strain, because F.prausnitzii has been reported as a hopeful next-generation probiotic against Crohn’s disease (4) .
Taking these MMODES simulated timeseries as input, we then use MDPbiome to model Crohn’s disease recovery.
MDPbiome can identify 3 microbiome states, which could be labelled as healthy, risky, and dysbiosis (F.prausnitzii ap-
prox. greater/equal/less than B.adolescentis, respectively). Fig.1B shows the state sequence of each simulated timeseries,
according to these state classiﬁcations. Given that higher Short Chain Fatty Acids concentration, such as butyrate, indi-
cates a healthier microbiome because it promotes bowel homeostasis, we sort the states maximizing the average production
of butyrate. Finally, the MDPbiome suggestion to recover from dysbiosis is to consume inulin as a prebiotic (as Fig. 1C
illustrates), with the model prediction that this will move the community to a healthier state, with a higher proportion
of F.prausnitzii than B.adolescentis, and a higher butyrate production.
Our system enables predictions of how diﬀerent amounts or types of ﬁber (diﬀerent chemical compositions, such
as starch, inulin or FOS) inﬂuence on the microbial community. Such analyses, with isolated ﬁber and at diﬀerent
concentrations, have not been undertaken experimentally (3), because food and even prebiotics mix diﬀerent types of ﬁber.
Thus, MDPbiomeGEM can assist with the planning and design of novel and health-relevant experiments of perturbed
We present a new approach to the simulation of microbial community longitudinal data under perturbations, using a
computational model of a Crohn’s disease microbial community based on GEM. Simulated data are then used as input
to MDPbiome, a system that recommends one or a sequence of perturbations to apply to a microbiome to guide it to a
healthier or more performant state. We ﬁnd the output model and policy from MDPbiome is consistent with both manual
examination of the simulated data, and with biomedical knowledge about inﬂuence of dietary ﬁber in Crohn’s disease
(3). This supports our proposal that the combination of GEM simulations, and subsequent analyses by MDPbiome, may
be an eﬀective way to generate predictive intervention plans (e.g. for personalized medicine) in the absence of suﬃcient
experimentally-derived metagenomics data.
Although there are limitations due to the lack of availability and minimal standardization and curation of the indi-
vidual strain GEMs, the main contributions of our MDPbiomeGEM approach are: 1) Simulate a variety of microbiome
perturbations, such as prebiotics changing the medium nutrients; probiotics to increase biomass; pathogens/antagonists,
such as adding a new strain to the community to compete for resources; 2) Further validation of the MDPbiome system
(engineering communities via perturbations); 3) Versatility in deﬁnition of microbiome states by biomasses or metabolic
ﬂuxes; 4) Analysis of growth dynamics; and 5) Enabling simulation taking host cells into account.
Modeling a microbiome as a combination of GEMs requires simplifying the microbial community into its most repre-
sentative species. Nevertheless, it allow us to model a speciﬁc biological phenomenon of interest, simulating the microbial
community dynamics under perturbations, and making it possible to suggest perturbations to modulate the microbiome
in desirable ways.
1. I. E. El-Semman et al., BMC Systems Biology 8, 41, issn: 1752-0509 (Apr. 2014).
2. D. Rios-Covian et al., FEMS Microbiology Letters 362, fnv176 (Sept. 2015).
3. C. Wong, P. J. Harris, L. R. Ferguson, International Journal of Molecular Sciences 17 (2016).
4. R. Martın et al., Frontiers in Microbiology 8(June 2017).
5. B. Garc´ıa-Jim´enez, T. De la Rosa, M. D. Wilkinson, Bioinformatics 34, i838–i847 (2018).
6. D. R. Garza, M. C. van Verk, M. A. Huynen, B. E. Dutilh, Nature Microbiology 3, 456–460 (Mar. 2018).
7. A. Succurro, D. Segr`e, O. Ebenh¨oh, mSystems 4(2019).