Statistical and Computational Methods for Microbiome Data Analysis

© 2021 Yiwen WangMicroorganisms are ubiquitous on earth, but our understanding of how microorganisms' function and composition relate to the environment remains limited. In this thesis, we refer to the microbiome as the collection of microorganisms and their genomes within a specific environment. In particular, we have focused on the analysis of microbiome profiled using 16S rRNA gene sequencing. Microbiome data have their inherent characteristics, which limit statistical analysis. These characteristics include a large number of zeroes, uneven library sizes, compositional and multivariate structures. Furthermore, as the microbial communities are very sensitive to their surrounding environment, microbiome data are prone to batch effects, i.e. any unwanted sources of variation that are unrelated to and obscure the biological factors of interest. Therefore, computational methods for batch effect management that take into account the data characteristics are needed. In the first part of this thesis, we reviewed existing methods to manage batch effects. These can be distinguished between correcting or accounting for batch effects. The former generates batch effect corrected data for any downstream analysis, while the latter includes batch effects as covariates in statistical models. However, methods correcting for batch effects do not consider the non-Gaussian distribution and multivariate structure of the data, while methods accounting for batch effects are limited to differential abundance analysis. We proposed a new method for batch effect correction based on Partial Least Squares Discriminant Analysis. "PLSDA-batch" is non-parametric and multivariate. It estimates latent components related to treatment and batch effects to remove batch variation whilst preserving treatment variation. Two variants were also developed to handle unbalanced batch x treatment designs and to select microbial variables that discriminate sample groups (such as treatments). We also implemented "PLSDA-batch" into an R package. In the second part of this thesis, we developed two frameworks for batch effect management including our proposed methods or variable selection in microbiome data analysis, which are available on GitHub (https://github.com/EvaYiwenWang). In the third part of this thesis, we presented the collaborative work on oral microbiome to identify microbial signatures discriminating between Rheumatoid Arthritis (RA) and healthy control subjects, thus uncovering new knowledge on the relationship between oral microbiome and RA disease