Fig. 1: Schematic of the MDSINE2 method for inferring interpretable dynamical systems models of microbiomes at scale.
From: Learning ecosystem-scale dynamics from microbiome data with MDSINE2
a, Input data are measurements of total bacterial concentration (for example, 16S rRNA gene quantitative polymerase chain reaction (qPCR)) and measurements of taxa abundances (for example, 16S rRNA gene amplicon sequencing). Measurements should be obtained from studies in which the microbiome undergoes perturbations, providing effective information for inference. b, MDSINE2 infers dynamical systems models from data with the option of automatically learning interaction modules, or groups of taxa that share the same interactions with other modules and perturbations. This is a more compact representation that is more readily interpretable than learning interactions among all microbes. c, Example microbial interaction networks for the same number of taxa without module learning (i) and with module learning (ii). d, MDSINE is fully Bayesian and propagates error throughout the model (i), providing estimates of uncertainty for all variables (for example, latent trajectory along with measurements and their uncertainty (ii), and indicator and interaction strengths for ecological interactions (iii)). Prob., probability. e–g, The software provides a variety of tools for analysis and visualization of the inferred dynamical system, including analyses of taxonomic composition and phylogeny of modules (e), formal analyses of ecosystem stability and interaction motifs (f), and keystoneness (quantitative impact on the ecosystem when modules are removed) (g).
