Fig 1.
Schematic overview for the BEEM-Static algorithm.
BEEM-Static takes relative abundances from a cross-sectional microbiome dataset (A), and runs an expectation-maximization algorithm (B) to estimate both biomass values (C), the interaction network (D) and carrying capacities (E).
Table 1.
Correlation methods tested.
Fig 2.
Benchmarking performance for network structure and edge directionality.
Note that CCREPE has two versions with Pearson and Spearman correlations (CCREPE.P and CCREPE.S), while SPIEC-EASI using “mb” and “glasso” algorithms is represented as SE.mb and SE.glasso, respectively. (A) ROC curves for different correlation based methods, the regression method using true biomass values and BEEM-Static for one simulated dataset with 30 species and 500 samples. (B) Boxplots showing precision and recall for directionality/sign of interactions for 30 different simulated communities with 30 species and 500 samples each. (C) True interaction network for a synthetic community based on all-pair co-culture data (Ground truth), inferred correlation network by BAnOCC and inferred interaction network by BEEM-Static (numbers on edges represent gLVM parameter values or correlation coefficients). PC: Prevotella copri, BV: Bacteroides vulgatus, BU: Bacteroides uniformis, BO: Bacteroides ovatus, BT: Bacteroides thetaiotaomicron, FP: Faecalibacterium prausnitzii, BH: Blautia hydrogenotrophica, ER: Eubacterium rectale, CA: Collinsella aerofaciens, EL: Eggerthella lenta, DP: Desulfovibrio piger, CH: Clostridium hiranonis.
Fig 3.
BEEM-Static robustly filters samples violating modeling assumption in simulated and real datasets.
(A-B) Performance reduction for BEEM-Static (with filters) and the naïve algorithm (without filtering of samples) as the percentage of samples at (A) equilibrium or (B) generated from the main model, decreases. Reduction is measured relative to BEEM-Static with no filters and all data from the model and at equilibrium. Points denote the means while error bars denote the standard deviation across 30 simulations each. (C-D) Principal coordinates plots (Bray-Curtis dissimilarity) representing gut microbiome taxonomic profiles from 4,617 samples. Points represent samples from individuals taking antibiotics (C) or from newborn infants (D), while crosses represent samples from adults who are not undergoing antibiotic treatment. Points that were filtered by BEEM-Static are colored blue and red otherwise.
Fig 4.
Analysis of human gut microbiomes with BEEM-Static.
(A) Violin plots showing the significant difference in BEEM-Static biomass estimates for adults and newborn infants. (B) Boxplots showing differences in DNA replication rates (GRiD estimates) for species predicted to decrease and increase in population size, respectively, by BEEM-Static. Each point represents one species in a sample. (C) Scatter plot showing the first two components from a principal component analysis of equilibrium abundances for samples (as predicted by BEEM-Static). (D) Predicted carrying capacities (square root transformed) of species in the two models. Species with divergent carrying capacities (ratio is >2 standard deviations from 1) in the two models are marked with stars.