Statistical signatures for adverse events in molecular life sciences

Abstract

The ongoing evolution of computational sciences is helping to address the growing data analytical needs in applications. For instance, in biosciences, recent advances in measurement technologies have resulted in large amounts of data with domain-specific properties that are challenging to analyze with traditional statistical methods.

An example of such a domain is microbiomics, the study of microbial communities, which in humans, have been reported to be associated with health and diseases. Despite advances in the field, further research is needed, as there is still a lack of understanding of how microbiome data should be processed and of the universal ecological properties of these complex systems.

The objective of this thesis is to advance the field of microbiome data science by considering methods for predicting future outcomes based on current information. This is achieved through developing time series methods for complex systems and applying established statistical models in large population cohorts.

The thesis consists of two complementary parts. The first part consists of analyses of two prospective human gut microbiome data sets, and contains the first ever microbiome-based survival analysis. The second part is focused on the stability properties of dynamical systems. It shows that the Bayesian statistical framework can be used to improve accuracy in inferring stability features, such as systemic resilience and early warning signals for catastrophic state transitions.

The results of this thesis contribute to the best practices of human microbiomerelated data science and demonstrate the advantages of the Bayesian framework in detecting adverse events in limited time series. Although the work was motivated by timely questions in microbiomics, the developed tools are generic and applicable in various contexts.