Statistical and network dynamics approaches to cancer genomics data analytics

In this thesis we focus on some statistical and physical methods which attempts to tackle the problem of cancer genetic heterogeneity and its relationship to higher level biological properties. The interactome allows to gain a system level view of mutational patters, providing a framework to understand how mutations act together to give rise to the cancer phenotype. Since different reconstruction of the interactome exist, in the first chapter of this thesis, we compare them from a topological perspective by analyzing their properties and we then study their overall resilience under node perturbation. Cancer stems from the impairment of one or more biological functions due to mutations of genes taking part in them. The observation that different patterns of mutations lead to different responses to treatments highlights the importance of stratifying patients based on their genetics and cytogenetic alterations. To this end, in the second chapter, we focus on hierarchical non parametric bayesian methods. Latent topic models allow to model hidden structures in the data and fit well with the hypothesis that cancer mutations impact specific gene groups in different proportions. In the second part of the chapter, we study a cohort of 2043 patients affected by Myelodysplastic Syndromes. From a more general perspective, the view of cancer as an evolutionary process, frequently implies the assumption of a direct and univocal genotype-phenotype relationship. However, as for cell differentiation, such genetic deterministic view is not always satisfactory. In the third chapter, we focus on the hypothesis of cancer as an abnormal attractor in the epigenetic landscape of the cell. We study the connection between the empirical distribution of cell in the gene expression state space with network laplacian-based manifold reconstruction techniques and their application for inferring the epigenetic landscape from data.