Identifying the molecular components that matter: a statistical modelling approach to linking functional genomics data to cell physiology

Functional genomics technologies, in which thousands of mRNAs, proteins, or metabolites can be measured in single experiments, have contributed to reshape biological investigations. One of the most important issues in the analysis of the generated large datasets is the selection of relatively small sub-sets of variables that are predictive of the physiological state of a cell or tissue. In this thesis, a truly multivariate variable selection framework using diverse functional genomics data has been developed, characterized, and tested. This framework has also been used to prove that it is possible to predict the physiological state of the tumour from the molecular state of adjacent normal cells. This allows us to identify novel genes involved in cell to cell communication. Then, using a network inference technique networks representing cell-cell communication in prostate cancer have been inferred. The analysis of these networks has revealed interesting properties that suggests a crucial role of directional signals in controlling the interplay between normal and tumour cell to cell communication. Experimental verification performed in our laboratory has provided evidence that one of the identified genes could be a novel tumour suppressor gene. In conclusion, the findings and methods reported in this thesis have contributed to further understanding of cell to cell interaction and multivariate variable selection not only by applying and extending previous work, but also by proposing novel approaches that can be applied to any functional genomics data.