Metabolite Overproduction through Engineering and Optimization of Microbiome Composition Dynamics

Although many advances have been made in genome sequencing for analyzing the composition of microbiomes, very few studies have attempted to learn and model their dynamics. Furthermore, no studies have attempted to exploit the dynamics of compositional changes of a microbiome for overproducing a metabolite of interest. This task proves to be computationally difficult at best and intractable at worst. The challenge is due to the complex, interdependent, and large number of highly non-linear interactions among members of a microbiome, as well as environmental factors, e.g. substrate. Here, we present a computationally tractable strategy using machine learning methods and stochastic optimization to characterize and potentially engineer a microbiome. In this work, an artificial neural network (ANN) is utilized to learn how six different lignocellulose food sources affect the temporal composition of the hindgut microbiome of Reticulitermes flavipes, the eastern subterranean termite. The learned dynamics from the ANN are optimized using either a genetic algorithm or artificial immune system approach. Specifically, the optimization objective is the maximization of the Rhodospirillales, an acetate producing order of bacteria, which will intrinsically maximize acetate production from the microbiome. The genetic algorithm and artificial immune system are compared for robustness and speed.