Predicting microbial interactions with approaches based on flux balance analysis: an evaluation.

BACKGROUND: Given a genome-scale metabolic model (GEM) of a microorganism and criteria for optimization, flux balance analysis (FBA) predicts the optimal growth rate and its corresponding flux distribution for a specific medium. FBA has been extended to microbial consortia and thus can be used to predict interactions by comparing in-silico growth rates for co- and monocultures. Although FBA-based methods for microbial interaction prediction are becoming popular, a systematic evaluation of their accuracy has not yet been performed. RESULTS: Here, we evaluate the accuracy of FBA-based predictions of human and mouse gut bacterial interactions using growth data from the literature. For this, we collected 26 GEMs from the semi-curated AGORA database as well as four previously published curated GEMs. We tested the accuracy of three tools (COMETS, Microbiome Modeling Toolbox and MICOM) by comparing growth rates predicted in mono- and co-culture to growth rates extracted from the literature and also investigated the impact of different tool settings and media. We found that except for curated GEMs, predicted growth rates and their ratios (i.e. interaction strengths) do not correlate with growth rates and interaction strengths obtained from in vitro data. CONCLUSIONS: Prediction of growth rates with FBA using semi-curated GEMs is currently not sufficiently accurate to predict interaction strengths reliably.

Deciphering Interactions Within a 4-Strain Riverine Bacterial Community.

The dynamics of a community of four planktonic bacterial strains isolated from river water was followed in R2 broth for 72 h in batch experiments. These strains were identified as Janthinobacterium sp., Brevundimonas sp., Flavobacterium sp. and Variovorax sp. 16S rRNA gene sequencing and flow cytometry analyses were combined to monitor the change in abundance of each individual strain in bi-cultures and quadri-culture. Two interaction networks were constructed that summarize the impact of the strains on each other's growth rate in exponential phase and carrying capacity in stationary phase. The networks agree on the absence of positive interactions but also show differences, implying that ecological interactions can be specific to particular growth phases. Janthinobacterium sp. was the fastest growing strain and dominated the co-cultures. However, its growth rate was negatively affected by the presence of other strains 10 to 100 times less abundant than Janthinobacterium sp. In general, we saw a positive correlation between growth rate and carrying capacity in this system. In addition, growth rate in monoculture was predictive of carrying capacity in co-culture. Taken together, our results highlight the necessity to take growth phases into account when measuring interactions within a microbial community. In addition, evidence that a minor strain can greatly influence the dynamics of a dominant one underlines the necessity to choose population models that do not assume a linear dependency of interaction strength to abundance of other species for accurate parameterization from such empirical data.

Technical versus biological variability in a synthetic human gut community.

Synthetic communities grown in well-controlled conditions are an important tool to decipher the mechanisms driving community dynamics. However, replicate time series of synthetic human gut communities in chemostats are rare, and it is thus still an open question to what extent stochasticity impacts gut community dynamics. Here, we address this question with a synthetic human gut bacterial community using an automated fermentation system that allows for a larger number of biological replicates. We collected six biological replicates for a community initially consisting of five common gut bacterial species that fill different metabolic niches. After an initial 12 hours in batch mode, we switched to chemostat mode and observed the community to stabilize after 2-3 days. Community profiling with 16S rRNA resulted in high variability across replicate vessels and high technical variability, while the variability across replicates was significantly lower for flow cytometric data. Both techniques agree on the decrease in the abundance of Bacteroides thetaiotaomicron, accompanied by an initial increase in Blautia hydrogenotrophica. These changes occurred together with reproducible metabolic shifts, namely a fast depletion of glucose and trehalose concentration in batch followed by a decrease in formic acid and pyruvic acid concentrations within the first 12 hours after the switch to chemostat mode. In conclusion, the observed variability in the synthetic bacterial human gut community, as assessed with 16S rRNA gene sequencing, is largely due to technical variability. The low variability seen in HPLC and flow cytometry data suggests a highly deterministic system.

Fast quantification of gut bacterial species in cocultures using flow cytometry and supervised classification.

A bottleneck for microbial community experiments with many samples and/or replicates is the fast quantification of individual taxon abundances, which is commonly achieved through sequencing marker genes such as the 16S rRNA gene. Here, we propose a new approach for high-throughput and high-quality enumeration of human gut bacteria in a defined community, combining flow cytometry and supervised classification to identify and quantify species mixed in silico and in defined communities in vitro. We identified species in a 5-species in silico community with an F1 score of 71%. In addition, we demonstrate in vitro that our method performs equally well or better than 16S rRNA gene sequencing in two-species cocultures and agrees with 16S rRNA gene sequencing data on the most abundant species in a four-species community. We found that shape and size differences alone are insufficient to distinguish species, and that it is thus necessary to exploit the multivariate nature of flow cytometry data. Finally, we observed that variability of flow cytometry data across replicates differs between gut bacterial species. In conclusion, the performance of supervised classification of gut species in flow cytometry data is species-dependent, but is for some combinations accurate enough to serve as a faster alternative to 16S rRNA gene sequencing.