Machine learning methods for the human microbiome

点击翻译按钮获取中文摘要

Therapeutic diets such as exclusive enteral nutrition for Crohn's disease can provide very effective treatments. However, they do not work for all individuals in all cases. The impact of these diets will be mediated through the gut microbiome, a diverse community of microbes involved in many aspects of gut metabolism but also inflammation through interaction with the immune system. The microbiome varies from one individual to another even in healthy individuals but exhibits substantially more variability in people suffering from IBD. It is highly likely therefore, that this variability impacts treatment efficacy.Cross-sectional studies have shown some organisms to be elevated or decreased in abundance in irritable bowel diseases (IBDs) such as Crohn's. However, these are purely associative observations. Longitudinal studies where patients are followed during treatment have far more statistical power. We will exploit a number of large-scale longitudinal data sets from dietary treatments of Crohn's disease that combine both metagenomics data which reveals the functional capacity of the microbiome together with metabolomics that measures small molecule concentrations produced either through the host or microbial metabolism. Coupling metabolomics with metagenomics has great potential to shift current microbiome research towards understanding community functions and interactions with the host. Previous studies analysed individual 'omics data, with many powerful bioinformatics tools developed over the past decade to enable metabolome and microbiome profiling. On the other hand, multi-omics data integration and interpretation is still a problem to be solved. The latter integration requires the deployment of advanced machine learning and statistical algorithms that leverage multiple heterogeneous, yet interconnected, data sources and that scale on large and high-dimensional datasets. Through the development and application of machine learning approaches we will quantify the importance of the initial host microbiome on treatment outcome but also the extent to which the metabolome under a controlled diet is determined by the patient microbiome. These machine learning methods will be combined with statistical modelling of the metabolic pathways in the gut, fitted to both 'omics data sources using approximate Bayesian methods on networks such as variational inference or expectation-propagation. From these we will infer which metabolic pathways are both microbially mediated and important for treatment. The development of these models will be aided by bench-top experiments generating longitudinal 'omics data sets from applying dietary perturbations to communities generated from patient and control fecal inocula in the artificial colon systems at the Quadram Institute.