Among the approaches used for modeling microbiomes, machine learning (ML) represents a powerful strategy. ML-based models are often able to overcome intractability issues associated with more mechanistic strategies, albeit at the cost of losing mechanistic knowledge. As a result, while ML models can be convenient, their “black box” nature can lead to potential hurdles with respect to engineering microbiome systems. Specifically, ML models are developed by providing input information to yield accurate output information. Engineers, however, often require a desired output to achieve a given level of performance or to meet a given specification. What is not known is the necessary input to achieve this output. Such a problem is referred to as an inverse function problem, where the ML model itself is the function that needs to be inverted. Evolutionary algorithms, which are class of stochastic optimization methodologies, are well suited to tackle such challenges. In this work, we used evolutionary algorithms to address the inverse function problem. We will present an approach to manipulate the composition of the termite microbiome composition to process lignocellulose substrates into a desired compound. Using experimental time course data of how the termite microbiome composition varied over 49 days on six different lignocellulosic food sources, we trained a deep learning network on composition dynamics. We then applied evolutionary algorithms to determine the feed strategy for maximization of acetate. Regardless of starting material on which a microbiome was fed, ultimately to maximize acetate, maple was predicted to be necessary as a substrate for the final six weeks.

Modeling and manipulation of microbial communities through the integration of machine learning and evolutionary algorithms

2019 Spring National Meeting

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.