Kanser biyolojisi için gen kümesi tabanlı sınıflandırma modelleri

As one of most prevalent and fatal diseases worldwide, cancer has been the focus of biomedical research for many decades. A wide range of different cancers with various causes have been identified and studied, and many treatment methods and drugs have been developed for these cancers. However, many open questions remain in understanding the mechanisms of cancer. With the advent of the high throughput sequencing technologies and the ever-increasing availability of genomic information gathered from cancer patients, researchers have successfully employed genomic information in diagnosis, prognosis and treatment of cancers. Still, given the large number of genes, their considerable correlation and the heterogeneity of cancers, genomic information alone often falls short of generating interpretable models for cancers. To alleviate these effects, pathways can be incorporated into models of cancer. In this thesis, using pathways, we have developed classification methods that produce interpretable models for progression of cancers and survival outcome. Patients usually experience the growth of a number of tumors in their life-time. However, not all these tumors pose a danger to the patient's life. To communicate the severity of the cancer, practitioners and researchers use the staging convention. The stage of a cancer encodes the level of its spread in the neighboring tissue and the body in general, where early-stages are assigned to local tumors and late-stages to the cancers that have spread in the body. Hence, understanding what drives cancers from early- to late-stages is an important question in understanding the mechanisms of cancer. The survival outlook of a patient is a similar measure for the severity of a cancer and understanding the mechanisms that affect the survival chances of a patient is immensely important in devising treatment strategies. In this work, we use multiple kernel learning for integrating pathways into the classification models for cancers. This allows for developing models that are more accurate and far more interpretable compared to models generated by conventional methods that do not use the pathway information. We then extend this method in several directions to improve accuracy and interpretability of the models. In the second part of this thesis, observing that the level of sparsity of the solutions can differ largely from caner to cancer and is often only indirectly affected by the parameters of the methods, we develop a model with an adjustable measure of sparsity and propose efficient solution methods for solving instances of this problem. In the third part of this thesis, given the known similarities between cancers, we develop a framework for building multitask classification models to improve the classification accuracy of cohorts with limited data using the similar cohorts with abundant data. Finally, we employ optimization techniques such as the cutting-plane method and Benders decomposition to improve the algorithmic performance of these methods to the point that they can be applied successfully to large-scale problems stemming from considering several cancers together in a multitask framework. The practical application of these methods is examined by applying them to the two aforementioned classification tasks across the Cancer Genome Atlas datasets for 27 cancer types. The results of these experiments indicate that the incorporation of pathways into these classification problems facilitates generating more interpretable and more accurate models, the similarity of cancers can be leveraged using a multitask framework towards the same goals, and the optimization methods developed here allow for solving these large-scale problems in efficient time using parallelization technologies.