BAYESIAN MACHINE LEARNING TOOLS FOR ANALYZING MICROBIOME DYNAMICS

The human microbiota plays an important role in health and disease, and its therapeutic manipulation is beingactively investigated for a wide range of diseases that span every NIH institute. Our microbiota are inherentlydynamic, and analyzing these time-dependent properties is key to robustly linking the microbiota to disease,and predicting the effects of therapies targeting the microbiota; indeed, longitudinal microbiome data is beingacquired with increasing frequency, and is a major component of many NIH-funded projects. However, there iscurrently a dearth of computational tools for analyzing microbiome time-series data, which presents severalspecial challenges including high measurement noise, irregular and sparse temporal sampling, and complexdependencies between variables. The objective of this proposal is to introduce new capabilities, improve on,and provide state-of-the-art implementations of tools for analyzing dynamics, or patterns of change inmicrobiome time-series data. The tools we develop will use Bayesian machine learning methods, which arewell-recognized for their strong conceptual and practical advantages, particularly in biomedical domains. Toolswill be rigorously tested and validated on synthetic and real human microbiome data, including publiclyavailable datasets and those from collaborators providing 16S rRNA sequencing, metagenomic, andmetabolomics data. We propose three specific aims. For Aim 1, we will develop integrated Bayesian machinelearning tools for predicting population dynamics of the microbiome and its responses to perturbations. Thesetools will include a new model that simultaneously learns groups of microbes with similar interaction structureand predicts their behavior over time, and that incorporates prior phylogenetic information. The model will befurther improved by incorporating stochastic microbial dynamics and errors in measurements throughout themodel. For Aim 2, we will develop Bayesian machine learning tools to predict host status from microbiomedynamics. The tools will learn easily interpretable, human-readable rules that predict host status frommicrobiome time-series data, and will be further extended to handle a variety of longitudinal study designs. ForAim 3, we will engineer our microbiome dynamics analysis software tools for optimal performance, ease-of-use, maintainability, extensibility, and dissemination to the community. In total, the proposed work will yield asuite of contemporary software tools for analyzing microbiome dynamics, with expected broad use and majorimpact. The software will allow investigators to answer important scientific and translational questions aboutthe microbiome, including discovering which microbial taxa or their metagenomes are affected over time byperturbations such as changes in diet or invasion by pathogens; predicting the effects of these perturbationsover time, including changes in composition or stability of the gut microbiota; and finding temporal signatures inmulti-?omic microbiome data that predict disease risk in the human host.