Retrieval of Gene Expression Measurements with Probabilistic Models

Measurement data sets of molecular biology and other experimental sciences are being collected comprehensively to openly accessible databases. This form of existing knowledge in quite complex, partly relevant and heterogeneous given that laboratories around the globe have different procedures to measure patient data. The research problem addressed in the thesis can be abstracted as "what can be done with the available large repositories towards cumulatively building knowledge from data in molecular biology". The thesis arrives at an answer: a "modelling-driven data retrieval engine", which researchers can use to position their measurement data into the context of earlier biology. Three types of commonly observed background repositories are considered as case studies and corresponding retrieval engines are proposed. In each case, the hidden relationships among the data are modelled and the modelled relationships, in turn, are used to compute relevance between two given data points. It is argued that using the available background information from hundreds of different situations or conditions in the proposed fashion allows to both complement the existing scarce data and to focus the analysis on relevant variables.

The main results of the thesis are the novel computational methods for the tasks of information retrieval and modelling hidden patterns in diverse data collections.