Dissecting patient heterogeneity via statistical modeling based on multi-modal omics data

Abstract

One of the key goals of modern medicine is to treat patients individually, recognizing the heterogeneity that exists within them and thus hoping to provide them with more effective personalized therapies. '-Omics' patient data provides a valuable resource to understand the patient heterogeneity and gain an insight into the biological phenomena at the intracellular level. As it is impossible to dissect patient groups based on single biomarkers or clinical factors, multivariate data mining and statistical modelling approaches (machine learning) play an important role. Moreover, understanding complex disease mechanisms calls for a more comprehensive and integrative approach, hence motivating the use of different kinds of data from the same patient. As individual -omics data sources capture specific kinds of molecular phenomena, there is a pressing need for multi-modal statistical approaches which combine several kinds of -omics data together.<br /> The present thesis addresses the aforementioned issues, viz. the exploration of heterogeneous patient populations based on their multi -omics profiles using statistical and machine learning approaches. More specifically, the main contributions of the thesis include: a) a retrospectively validated prediction model for GBM (Glioblastoma Multiforme) recurrence location and b) development of a new algorithm- Survival based Bayesian Clustering which is a merger of clustering and supervised prediction, this algorithm has been successfully shown to be an important step towards the discovery of clinically relevant patient strata leveraging the potential of multi-omics data integration. <br /> The novel algorithm of Survival based Bayesian Clustering was tested successfully in various scenarios and on different patient populations.The thesis also provides a deep understanding of our proposed technique from a purely statistical point of view. Overall, work in this thesis is a step forward in moving towards the goal of personalized medicine solutions using multi-modal molecular -omics data and statistical modelling.