Prof. Jaak Vilo (Institute of Computer Science UT)
senior researcher Hedi Peterson (Institute of Computer Science UT)
Prof. Laura Elo (University of Turku, Finland)
Prof. Ewa Szczurek (University of Warsaw, Poland)
Summary of the Thesis:
A fast advance in biotechnological innovation and decreasing production costs led to explosion of experimental data being produced in laboratories around the world. Individual experiments allow to understand biological processes, e.g. diseases, from different angles. However, in order to get a systematic view on disease it is necessary to combine these heterogeneous data. The large amounts of diverse data requires building machine learning models that can help, e.g. to identify which genes are related to disease. Additionally, there is a need to compose reliable integrated data sets that researchers could effectively work with. In this thesis we demonstrate how to combine and analyze different types of biological data in the example of three biological domains: Alzheimer’s disease, immunology, and toxicology. More specifically, we combine data sets related to Alzheimer’s disease into a novel heterogeneous network-based data set for Alzheimer’s disease (HENA). We then apply graph convolutional networks, state-of-the-art deep learning methods, to node classification task in HENA to find genes that are potentially associated with the disease. Combining patient’s data related to immune disease helps to uncover its pathological mechanisms and to find better treatments in the future. We analyse laboratory data from patients’ skin and blood samples by combining them with clinical information. Subsequently, we bring together the results of individual analyses using available domain knowledge to form a more systematic view on the disease pathogenesis. Toxicity testing is the process of defining harmful effects of the substances for the living organisms. One of its applications is safety assessment of drugs or other chemicals for a human organism. In this work we identify groups of toxicants that have similar mechanism of actions. Additionally, we develop a classification model that allows to assess toxic actions of unknown compounds.