Fig. 2

Overview of MC-TIMME2 computational model and representative individual taxa trajectories inferred from the model. a MC-TIMME2 is an unsupervised Bayesian nonparametric machine learning method that takes as input time-series data of microbial counts (e.g., tables of counts derived from shotgun metagenomics sequencing), dosing of the compound for each participant, and a microbial phylogenetic tree. Two levels of clustering are simultaneously learned to: (1) discover groups of common kinetic parameters (microbe groups, incorporating phylogenetic information) and common responses to the compound (perturbation groups). Microbe groups are characterized by common growth rate and carrying capacity parameters. Perturbation groups are characterized by a time-period of activity (including possible time-delays) and a magnitude of the perturbation, modulated by inferred participant-specific compound levels passed through a nonlinear transfer function. Microbe groups are super-clustered into perturbation groups. MC-TIMME2 produces several outputs, including carrying capacities (estimated steady-state levels on and off perturbations), inferred participant-specific compound concentrations over time, and maps of perturbation response clusters. b Trajectories of individual taxa inferred by MC-TIMME2, selected to highlight model capabilities. Ruminococcaceae bacterium cv2 in subject 201-021 shows a strong positive perturbation effect that ends before the end of the feeding period. Bifidobacterium longum and Bifidobacterium adolescentis in subject 101-018 demonstrate a delayed positive response that does not start until about 10 days after the start of compound administration. Tyzzerella nexilis in subject 201-024 shows a strong negative response. Clostridium leptum in subject 201-028 shows an increase in the magnitude of the response as the dose of the compound was increased