CODY enables quantitatively spatiotemporal predictions on in vivo gut microbial variability induced by diet intervention

Microbial variations in the human gut are harbored in temporal and spatial heterogeneity, and quantitative prediction of spatiotemporal dynamic changes in the gut microbiota is imperative for development of tailored microbiome-directed therapeutics treatments, e.g. precision nutrition. Given the high-degree complexity of microbial variations, subject to the dynamic interactions among host, microbial, and environmental factors, identifying how microbiota colonize in the gut represents an important challenge. Here we present COmputing the DYnamics of microbiota (CODY), a multiscale framework that integrates species-level modeling of microbial dynamics and ecosystem-level interactions into a mathematical model that characterizes spatial-specific in vivo microbial residence in the colon as impacted by host physiology. The framework quantifies spatiotemporal resolution of microbial variations on species-level abundance profiles across site-specific colon regions and in feces, independent of a priori knowledge. We demonstrated the effectiveness of CODY using cross-sectional data from two longitudinal metagenomics studies—the microbiota development during early infancy and during short-term diet intervention of obese adults. For each cohort, CODY correctly predicts the microbial variations in response to diet intervention, as validated by available metagenomics and metabolomics data. Model simulations provide insight into the biogeographical heterogeneity among lumen, mucus, and feces, which provides insight into how host physical forces and spatial structure are shaping microbial structure and functionality.

Changes in the human gut microbiome composition are connected with development of numerous diseases, like obesity, type-2 diabetes, and immune dysfunction (1–3). Quantitative understanding and predicting how microbial variations are determined are crucial for designing microbiome-directed therapies that target chronic metabolic diseases (4, 5). However, this remains challenging due to the temporal and spatial heterogeneity along the human gut resulting from a dynamic interplay among host, microbial, and environmental conditions (6, 7). Diet is recognized as a controllable and pivotal environmental factor in shaping longitudinal microbial landscape development (8, 9), such as early childhood colonization (10) and long-term adulthood stabilization (11). While profiling of fecal samples enables a snapshot of consequential changes of the fecal microbiota in response to different stimuli, e.g. dietary changes (12–14), it is still far from describing the intrinsic dynamics of how microbiome colonize in the gut. Recently, the spatial heterogeneity of microbial composition between lumen and mucus has been recognized in mice (15, 16), but similar studies in humans is impossible with current techniques. In addition, measurements of absolute abundance profiles are required to correct the artifacts associated with relative abundance that confound revealing the interplay between microbial variations and health (17). Therefore, methods that enable quantifying the absolute, temporal, and spatial variations of in vivo human gut microbiota are needed to understand how to maintain or restore healthy microbiota.Computational models are widely used to decipher microbial complexity and response to perturbations (18). Most existing models have limited usage as they only address specific elements of the multidimensional interaction mechanisms. For example, similarity-based (19) and rule-mining models (20) describe microbial–microbial interactions without considering temporal dependency. The dynamic Bayesian model enables incorporation of directed interactions and longitudinal dataset (21), while reliance on training dataset and difficulties in model selection render these stochastic models confining to specific statistic condition and predictions are therefore not consistent and generalizable (22). The generalized Lotka–Volterra model (18, 23, 24) represents a step forward to simulate dynamics via formulating microbial growth rate as a lumped term, but adherence to assumptions of pairwise interactions-driven community dynamics and constant environment limits their predictive power. Genome-scale models (GEMs) (25) provide a valuable resource for studying structured microbial metabolism. With GEMs, microbial metabolic capacity, microbe–microbe interactions (26–28), microbial–diet interactions (12), and structural changes of two-species cocultures (29) are characterized using flux balance analysis (FBA). With rare exceptions, FBA requires a priori knowledge of metabolite uptake fluxes distributed among community members, with current limitations on these resources, faces challenges with modeling multispecies communities in a dynamic manner (30). Therefore, in adapting a computational framework that can simulate microbiome dynamics along the human gut, one encounters three challenges: 1) endogenously, the intrinsic dynamics not only emerge from the large number of microbiota components but also through the intricate and dynamic ways they interact (31, 32); 2) exogenously, the microbiota is exposed to a series of host–microbe metabolic axes (33), such as colonic physical forces (34), nascent colonization, and nutrient conditions; and 3) spatial structure of the in vivo microbiota localization plays a significant role impacting 1 and 2 (24).Here, we bridge the current theoretical gap by developing a multiscale framework for COmputing the DYnamics of gut microbiota (CODY), which enables identification and quantification of spatiotemporal-specific variations of gut microbiome absolute and relative abundance profiles, without a prior knowledge of microbiome interactions. We evaluated CODY’s performance by comparing model simulations with longitudinal changes in the microbial composition in fecal samples and in plasma metabolomics of two cohorts: 1) long-term development of the gut microbiome in early infancy and 2) short-term variation patterns of the gut microbiome in obese adults experiencing diet intervention. Comparison of model simulations with experimental data demonstrated predictive strength of the CODY modeling framework and hence lays the foundation for performing design of microbiome-directed therapeutics or of precision nutrition based on CODY simulations. The source code of CODY is freely available together with full documentation at https://github.com/JunGeng-Sysbio-Chalmers/CODY1.0_SourceCode.