Novel methods for inferring species interaction strengths from microbial time series data

Background/Question/Methods

One method of modeling the dynamics of a microbial community is through generalized Lotka-Volterra equations.  In this approach, it is assumed that pairwise interactions, along with growth and intraspecific competition, are the main factors controlling growth dynamics for each species in the community.  If robust and reliable, the impacts to quantifying interspecific relationships in this way can help analyze and predict the prevalence of certain helpful or harmful species in the human microbiome and other microbial communities. We quantify community structure in the form of an interaction matrix, where each value represents the specific effect of a given species on the abundance of another species.  Recent studies have used metagenomic time-series sequence data from microbial communities to infer the strength of these interactions. This type of inference is difficult in part because the metagenomic data does not show absolute species abundances, and also because the available data can easily lead to overfitting. 

Results/Conclusions

We test two inference algorithms drawn from the literature, each of which uses a different approximation to infer maximum likelihood parameters for interaction matrices. We also introduce our own new inference algorithm, based on approximating likelihoods using moment closure methods. We compare the results of these three approaches against simulated, stochastic community data, both for communities with randomly generated interaction strengths, and for communities whose interactions have been allowed to evolve over time, and thus have non-random community interaction structure. We find that our algorithm outperforms existing approaches, quantified using absolute parameter values, predicted species abundances, and the eigenvalue spectrum of inferred matrices of interactions. The differences are most pronounced for the `realistic’ communities generated by allowing interactions to evolve over time. Our final step applies all three algorithms to publicly-available microbiome time-series data and we discuss the most significant differences in inferred interaction strengths.