Association, causality, and biomarker discovery in translational microbiome research

Abstract

This dissertation develops statistical methods for biomarker detection in high-dimensional microbiome studies. The human microbiota is essential for health, and disruptions in its composition are linked to many diseases. Advances in sequencing technologies have enabled the collection of large-scale microbiome data, creating both opportunities and methodological challenges for identifying biomarkers. The focus is on microbiome intervention studies, where intervention factors are administered to subjects, followed by measurements of the microbiome and clinical outcomes. The goal is to investigate the relationships among intervention factors, microbiome features, and clinical outcomes, and to identify microbiome biomarkers while accounting for intervention effects. First, feature-specific modelling approaches are considered, where individual microbial features are evaluated as potential biomarkers. Building on existing frameworks, a structural equation modelling approach is proposed to provide a more detailed characterization of the relationships between biomarkers and outcomes by decomposing the underlying association structures. Next, the limitations of single-feature models are addressed by introducing methods for constructing multiple biomarkers based on sets of features. These approaches use penalized regression techniques to perform simultaneous feature selection and prediction in high-dimensional settings, with Monte Carlo cross-validation applied to reduce bias and overfitting. The framework is flexible and accommodates different types of outcomes, including continuous, binary, and time-to-event responses. Although motivated by microbiome intervention studies, the proposed methods are broadly applicable to other high-dimensional data. Overall, this work contributes modelling strategies for identifying microbiome biomarkers for clinical outcomes while adjusting for the effects of intervention factors.