An expectation-maximization algorithm enables accurate ecological modeling using longitudinal microbiome sequencing data

Abstract

Background The dynamics of microbial communities is driven by a range of interactions from symbiosis to predator-prey relationships, the majority of which are poorly understood. With the increasing availability of high-throughput microbiome taxonomic profiling data, it is now conceivable to directly learn the ecological models that explicitly define microbial interactions and explain community dynamics. The applicability of these approaches is severely limited by the lack of accurate absolute cell density measurements (biomass). Methods We present a new computational approach that resolves this key limitation in the inference of generalized Lotka-Volterra models (gLVMs) by coupling biomass estimation and model inference with an expectation-maximization algorithm (BEEM). Results BEEM outperforms the state-of-the-art methods for inferring gLVMs, while simultaneously eliminating the need for additional experimental biomass data as input. BEEM’s application to previously inaccessible public datasets (due to the lack of biomass data) allowed us to construct ecological models of microbial communities in the human gut on a per-individual basis, revealing personalized dynamics and keystone species. Conclusions BEEM addresses a key bottleneck in “systems analysis” of microbiomes by enabling accurate inference of ecological models from high throughput sequencing data without the need for experimental biomass measurements. Electronic supplementary material The online version of this article (10.1186/s40168-019-0729-z) contains supplementary material, which is available to authorized users.

BEEM-Static: Accurate inference of ecological interactions from cross-sectional metagenomic data

“…We anticipate that the incorporation of such environmental factors into future models would be an exciting avenue to study their influence on microbial community structure in vivo. Finally, hybrid methods that learn models from both longitudinal and cross-sectional data represent another promising direction to explore for studying general and individual specific microbiome dynamics (Li et al, 2019). Biomass estimates…”

“…Here we show that an expectation maximization algorithm which couples gLVM parameter inference with scaling factor estimation (BEEM -originally designed for longitudinal data [Li et al, 2019]) can be transformed to work with cross-sectional data from communities that are at or near equilibrium (BEEM-Static). In benchmarking comparisons with simulated communities against 10 other methods that infer microbial interactions from cross-sectional data, we noted that while all other methods only improved slightly over random predictions (AUC-ROC<0.63), BEEM-Static exhibited high accuracy similar to estimation using true scaling values (AUC-ROC>0.88).…”

From bag-of-genes to bag-of-genomes: metabolic modelling of communities in the era of metagenome-assembled genomes

“…This suggests that the fate of the communities in terms of composition was mainly driven by the availability of nutrients, which can be explained by a generic consumer-resource model. In addition, gLV models have been extensively applied to study the temporal evolution of communities [119] , [137] , [138] , [139] , [140] . These models enable predictions of community dynamics by taking into account growth rates and interactions strengths between microbes [141] .…”

Compositional Lotka-Volterra describes microbial dynamics in the simplex

“…However, these methods implicitly ignore the constraint that relative abundances must sum to one and are therefore negatively correlated, making parameter estimates difficult to interpret. Li et al [34] suggest addressing this by inference of the latent overall biomass. Alternatively, Shenhav et al [21] suggested a linear mixed model with variance components, while representing the previous state microbial community using its quantiles instead of relative abundances.…”