UC San Diego

Next-generation sequencing (NGS) has effected an explosion of research into the relationship between genetic information and a variety of biological conditions. One of the most exciting areas of study is how the trillions of microbial species that we share this Earth with affect our health. However, the process of extracting useful biological insights from this breadth of data is far from trivial. There are numerous statistical and computational considerations in addition to the already complex and messy biological problems. In this thesis, I describe my work on developing and implementing software to tackle the complex world of statistical microbiome analysis.In the first part of this thesis, we review the applications and challenges of performing dimensionality reduction on microbiome data comprising thousands of microbial taxa. When dealing with this high dimensionality, it is imperative to be able to get an overview of the community structure in a lower dimensional space that can be both visualized and interpreted. We review the statistical considerations for dimensionality reduction and the existing tools and algorithms that can and cannot address them. This includes discussions about sparsity, compositionality, and phylogenetic signal. We also make recommendations about tools and algorithms to consider for different use-cases. In the second part of this thesis, we present a new software, Evident, designed to assist researchers with statistical analysis of microbiome effect sizes and power analysis. Effect sizes of statistical tests are not widely reported in microbiome datasets, limiting the interpretability of community differences such as alpha and beta diversity. As more large microbiome studies are produced, researchers have the opportunity to mine existing datasets to get a sense of the effect size for different biological conditions. These, in turn, can be used to perform power analysis prior to designing an experiment, allowing researchers to better allocate resources. We show how Evident is scalable to dozens of datasets and provides easy calculation and exploration of effect sizes and power analysis from existing data. In the third part of this thesis, we describe a novel investigation into the joint microbiome and metabolome axis in colorectal cancer. In most cases of sporadic colorectal cancers (CRC), tumorigenesis is a multistep process driven by genomic alterations in concert with dietary influences. In addition, mounting evidence has implicated the gut microbiome as an effector in the development and progression of CRC. While large meta-analyses have provided mechanistic insight into disease progression in CRC patients, study heterogeneity has limited causal associations. To address this limitation, multi-omics studies on genetically controlled cohorts of mice were performed to distinguish genetic and dietary influences. Diet was identified as the major driver of microbial and metabolomic differences, with reductions in alpha diversity and widespread changes in cecal metabolites seen in HFD-fed mice. Similarly, the levels of non-classic amino acid conjugated forms of the bile acid cholic acid (AA-CAs) increased with HFD. We show that these AA-CAs signal through the nuclear receptor FXR and membrane receptor TGR5 to functionally impact intestinal stem cell growth. In addition, the poor intestinal permeability of these AA-CAs supports their localization in the gut. Moreover, two cryptic microbial strains, Ileibacterium valens and Ruminococcus gnavus, were shown to have the capacity to synthesize these AA-CAs. This multi-omics dataset from CRC mouse models supports diet-induced shifts in the microbiome and metabolome in disease progression with potential utility in directing future diagnostic and therapeutic developments. In the fourth chapter, we demonstrate a new framework for performing differential abundance analysis using customized statistical modeling. As we learn more and more about the relationship between the microbiome and biological conditions, experimental protocols are becoming more and more complex. For example, meta-analyses, interventions, longitudinal studies, etc. are being used to better understand the dynamic nature of the microbiome. However, statistical methods to analyze these relationships are lacking – especially in the field of differential abundance. Finding biomarkers associated with conditions of interest must be performed with statistical care when dealing with these kinds of experimental designs. We present BIRDMAn, a software package integrating probabilistic programming with Stan to build custom models for analyzing microbiome data. We show that, on both simulated and real datasets, BIRDMAn is able to extract novel biological signals that are missed by existing methods. These chapters, taken together, advance our knowledge of statistical analysis of microbiome data and provide tools and references for researchers looking to perform analysis on their own data.