UNDERSTANDING DIFFERENTIAL ABUNDANCE IN MICROBIAL ECOLOGY USING COMMUNITY STRUCTURE.

Advances in DNA sequencing have deepened our understanding of the microbiome’s role in health and disease, with microbiome research spanning biotechnology, food science, agriculture, and medicine. Differential abundance analysis (DAA) is one of the central statistical methods for identifying microbial features that are associated with macroscopic states. Differentially abundant microbial features can broaden the understanding of disease mechanisms and guide prevention, diagnosis, prognosis, and therapy, yet current DAA methods often yield inconsistent, irreproducible results and prioritise isolated feature classification over the complex interaction networks. This thesis introduces a novel framework that reframes DAA through probabilistic graphical model inference, integrating network analysis with abundance profiling to uncover not only key microbial features but also the intricate interactions within their communities. Central to this work is the development of q2-Makarsa, a QIIME 2 plugin that provides a pipeline necessary to perform DAA using interaction network inference. q2-Makarsa exposes SpiecEasi and FlashWeave for the purpose of network inference, including the use of metavariables to represent exogenous influences on the microbiome, automates the analysis of inferred networks for the purpose of DAA, and provides publication-quality visualisations of the microbial networks. It also conveniently exposes consensus Louvain community detection to aid in network understanding. This enriched version of DAA goes beyond conventional classification of features as differentially abundant to reframe the question as one of influence. In benchmarking studies, where the performance of q2-Makarsa for DAA was compared against established methods such as ANCOM-BC and ANCOM-BC2, it consistently demonstrated superior performance in F1 scores when simulating using a third-party microbial abundance data simulator. In addition, a custom data simulator was developed to generate synthetic datasets that preserve the network structure of real-world data. The benchmark study using this new data simulator also shows promising performance. The results also indicate that the simulator may overemphasise network-driven signal propagation relative to other biological sources of variation. This work represents a significant contribution in microbial ecology and bioinformatics by merging network theory with DAA, offering a more comprehensive tool for disease diagnosis, prognosis, and treatment selection.