Kanser hastalığında önemli gen kümelerini belirlemek için geliştirilen en iyileme modelleri

Using genomic characterizations of tumours biopsied from cancer patients has a great importance in understanding the formation and progression mechanisms in cancer. Survival analysis is one of the research methods that is used to predict overall survival time of cancer patients and to understand the aforementioned progression mechanisms. High dimensional structure of the genomic characterizations with the limited number of training samples makes survival analysis a challenging task. To be able to identify the survival associated biological mechanisms, cancer-specific pathway/gene set collections can be integrated into machine learning models. Existing approaches usually follow a two-stage approach that either identify predictive genes using a feature selection method and map these selected genes to known pathways/gene sets, or train separate models for each pathway/gene set and try to pick informative ones considering each model's predictive performance. Following such a two-stage approach might result in inefficacy of mapping selected genes to a known biological pathway/gene set due to highly correlated structure between feature groups or including related or very similar pathways/gene sets into the final model due to analyzing each pathway/gene set separately. In this thesis, rather than following such two-stage approaches, we propose machine learning models that can conjointly identify disease related biological mechanisms and perform survival prediction using only these identified biological mechanisms. Our algorithms obtain a sparse set of pathways/gene sets for the survival associated biological mechanisms by eliminating the uninformative ones from the model. We test our algorithms using 20 cancer datasets obtained from The Cancer Genome Atlas and two cancer-specific pathway/gene set collections as input data. We first propose a survival analysis model that integrates pathway/gene set collection into the model using multiple kernel learning. Our algorithm with conjoint modelling approach obtains statistically significantly better or comparable predictive performances against survival random forest (RF) and survival support vector machine (SVM) using significantly fewer gene expression features. Predictive performances of machine learning algorithms can be increased using multitask learning. For this purpose, we extend our multiple kernel learning-based algorithm towards multitask learning. Our multitask learning algorithm both models multiple cancer datasets simultaneously and integrates cancer related biological mechanisms into the machine learning model. The algorithm is able to identify common underlying biological mechanisms for cancer by obtaining better or comparable predictive results against survival RF, survival SVM, and our multiple kernel learning survival analysis algorithm. We also extend our multitask learning algorithm towards task clustering to identify the groups of cancer types that share similar underlying biological mechanisms. To this aim, we propose a unified formulation for task clustering, survival analysis, and knowledge extraction. Our clustering algorithm identifies relevant cancer groups by obtaining statistically significantly better or comparable predictive performances against survival RF, survival SVM, our multiple kernel learning and multitask multiple kernel learning survival analysis algorithms. Numbers of gene expression features and gene sets used by our clustering algorithm are significantly fewer than those of benchmark algorithms. These results show that our methods that identify survival associated biological mechanisms, obtain better or comparable predictive performances against survival analysis methods developed on genomic data without using the pathway/gene set information in the literature. In addition, we prove that survival prediction can be performed using fewer number of gene expression features compared to these benchmark algorithms. We also identify the cancer groups that share similar biological mechanisms without decreasing the predictive power.