Abstract:
|
We present MDSINE, a Bayesian framework for inferring dynamical systems models from microbiome time-series datasets and predicting future behaviors of the microbiota. MDSINE uses a two-stage procedure for fast approximate inference of the posterior probability distribution of the underlying dynamical system's parameters. A posterior distribution over time-dependent bacterial growth trajectories and gradients is first estimated from count-based and continuous-valued molecular data using a novel Bayesian adaptive spline model. Statistics from this distribution are then used to infer parameters of an approximated nonlinear ordinary differential equations model, using priors to encourage sparse connectivity in the dynamical systems model. Using data simulated to mimic key properties of real microbiome studies, we demonstrate that MDSINE significantly outperforms the existing method for microbial dynamical systems inference on multiple metrics. We then demonstrate the utility of our method on two new gnotobiotic mice experimental datasets, investigating infection with an enteric pathogen, Clostridium difficile, and stability of an immunomodulatory probiotic cocktail.
|