We analyzed the artificial gut dataset described below using a MALLARD model that is generative and assumes there exists an unobserved microbial composition ( " ; the state) that evolves through time ( Fig. 1A) due to stochastic biological variations ( " ). 115We regard the state sequence ( $ , … , " , … , ' ) as the true microbial dynamics in a timeseries. Random technical variations ( " ) are then added to the true system state ( " ) resulting in the composition " (Fig. 1B). We observe " through a multinomial counting process. This formulation is similar to the constant level model commonly used in Bayesian time-series analysis (34). By separately modeling the process generating " 120 and " with distinct covariance matrices ( and , respectively), we can decouple biological and technical variations in artificial gut datasets (Fig. 1C). Visually, we found this model provided a good fit to artificial gut data (Fig. S1).
Daily and hourly gut microbiota time-series in an artificial human gut 125We applied our model to an artificial gut that was constructed using continuousflow anaerobic bioreactor systems that have been validated as models of human gut microbiota (1,6,15, 35,36). The same starting human fecal inoculum was seeded into replicate ex vivo vessels (n=4) and cultured for 1 month (Fig. 2 and S2). Throughout the experiment pH, temperature, media input rates, and oxygen concentration were all 130 fixed (Methods). To introduce microbial dynamics into our systems, a single bolus of Bacteroides ovatus isolated from the stool donor was administered to the system on Day 23 (Methods). The B. ovatus bolus did not have discernable effects on microbial dynamics, but the media it was suspended in appeared to induce minor shifts in the relative abundances of select bacterial taxa that were visible with hourly sampling (Fig. 135