Statistical approaches for metabolomics and omics data integration

Abstract

Biological processes are the result of multiple interactions between various omic entities and are inherently complex. Metabolomics profiling plays a key role into deciphering mechanisms of biological functions in living organisms and is hence gaining popularity. In the last twenty years, the parallel acquisition of high-throughput datasets from the genome, metabolome, proteome, and transcriptome has seen a tremendous boost. The integrative analysis of these datasets is promising to enhance the understanding of biological functions and uncover their underlying mechanisms. The main objectives of this thesis consist in i) developing and investigating novel statistical models for integrative analysis of metabolomics data with other omics technologies, ii) enriching the offer of probabilistic models tailored to metabolomics data and iii) providing enhanced interpretability of results. This thesis is mainly concerned with designing models that can be introduced in different steps of a typical analysis pipeline of metabolomics data. Chapters 2 and 3 motivate the importance of data integration and review popular statistical techniques used in metabolomics. Chapter 4 simultaneously covers two steps of the analysis pipeline, by building a single integrative Bayesian model that is able to perform both cross-omics biomarker discovery and infer potential perturbed pathways. Chapter 5 focuses solely on integrative statistical analysis by uncovering hidden associations between multi-omics data. Finally, in Chapter 6 we investigate the incorporation of pathway information into a Bayesian nonparametric clustering model and its potential to help metabolite annotation. Where possible, simulation studies are used to get a better understanding of our methods and test their applicability. These simulations are always followed by analysis of real data and comparison to competing methods. In most instances, our methods have resulted in plausible biological findings when applied to real data, and represent, to our knowledge, one of the first applications of such probabilistic models in integrative analysis of metabolomics data.