Deep Learning Frameworks for Multi-omics Analyses of the Microbiome in Disease Studies
thesisposted on 01.12.2021, 00:00 authored by Derek Reiman
Over the last decade, our understanding of the functions of the microbiome has greatly increased. In particular, the gut microbiome has been shown to play a critical role in the development of multiple metabolic diseases. As such, clinicians have looked to the microbiome community as a diagnostic marker as well as a potential target for therapeutic interventions, either by directly altering the microbiome composition or by targeting underlying metabolic functions. However, the development of these therapies is challenging. It requires the identification of which microbes are associated with the disease, an understanding of the underlying function of these microbes contributing to metabolic dysregulation, and the knowledge of how specific patient characteristics and stimuli can alter the dynamics of the microbiome community. In this dissertation, I will present novel deep learning frameworks that I have developed to address these challenges in microbiome studies. First, I will introduce a deep learning framework that integrates phylogenetic information with microbiome abundance to predict host disease status. I will demonstrate the predictive power of this model using complex disease states across Inflammatory Bowel Disease, Type 2 Diabetes, and other diseases, and show its ability to identify disease-related microbes at different taxonomic levels. Next, I will present an interpretable deep neural network framework integrating microbiome and metabolome data to predict the entire metabolomic profile from microbiome abundance. I will then show this framework can cluster microbes and metabolites into functionally related modules in order to identify underlying microbe-metabolite interactions. Lastly, I will introduce a deep learning framework for modeling the dynamics of the microbiome community while considering multiple patient characteristics and external factors. This framework not only improves the prediction performance compared to current state-of-the-art methods, but also facilitates the identification of host characteristics and factors that have a significant impact on microbiome dynamics. Together, these methods aim to provide a suite of robust and scalable tools to assist researchers in identifying microbes related to disease, uncovering the metabolic function of these microbes, and identifying potential treatment routes for improving patient health through microbiome targeted therapeutics.