Extended Data Fig. 6: Other approaches to predict assembly composition performed worse than the coarse-grained consumer-resource model based on metabolomics and pairwise spent-media experiments.

(a) Hypothetical resource utilization structures. The ‘base’ structure was defined as the set of species-specific resource groups. On top of the base structure, pairwise niche overlaps consumed by only two species and all-but-one niche overlaps consumed by 14 of the 15 species were also tested. (b) Performance of utilization structures selected by regularized regression on all detected resource groups (Methods). Shown are mean errors and coefficients of determination for LASSO fits. Shading denotes standard error of the mean. (c) Prediction errors of the full model as in Fig. 3a (left) versus model predictions after randomly shuffling species identity (right). Shown are box plots denoting the mean error (thick central mark), the 25th and 75th percentiles (box), and the extremes (dashed lines) across all assemblies tested (n = 185 assemblies). (d) The consumer-resource model achieved comparable performance as a Lotka-Volterra model fitted to all assembly data. Shown are box plots as in (c) across pairwise co-cultures (n = 105), assemblies with more than two species (n = 80), and all assemblies (n = 185 assemblies). Colors denote different models: the consumer-resource model (orange); and Lotka-Volterra models parametrized using pairwise spent-media experiments (black), species abundances in pairwise co-cultures (dark gray), or species abundances in all assemblies (light gray). (e) Lotka-Volterra models parametrized using assembly data failed to predict yield in pairwise spent-media experiments. (f) The coarse-grained consumer-resource model was the best performing model for both the mean absolute error of log₂(fold-change) (Fig. 3b) and the commonly used Bray-Curtis dissimilarity metric, defined as \(1-{\sum }_{i=1}^{N}\min \left({x\,}_{i}^{\text{actual}},{x\,}_{i}^{\text{predicted}}\right)\). Shown are box plots as in (c) across all assemblies tested (n = 185 assemblies) for the same models as in Fig. 3b. (g) The model successfully predicted absolute abundances, obtained by multiplying relative abundances by culture yield in OD. Panels are representative assemblies, analogous to Fig. 3c.

Source data