Çok katmanlı perceptron kullanarak erken ve geç evre kanserleri ayırma

Prediction of cancer stages is essential to early diagnosis of the tumours and to develop appropriate therapeutic strategies for cancer patients. Determining the biomarker genes is extremely important for understanding the progression of cancer from early-stages to late-stages. Standard machine learning algorithms could be adopted to build classifiers based on genomic characterizations to discriminate between early- and late-stage cancers. Some of these algorithms have limited knowledge extraction capability due to the strongly correlated nature of high-dimensional genomic data even if they have satisfactory prediction accuracies. In this thesis, to identify the relevant molecular mechanisms of the biological processes which might be driving cancer progression, we implemented an algorithm that integrates pathways/gene sets into our neural network model. We developed a multilayer perceptron (MLP) method on multiple pathways/genes to separate early- and late-stage of cancers. Our main goal was to determine biomarkers that have an effect on the progression of cancer and to obtain comparable predictive performance. We compared the performance of our MLP model with three existing classification algorithms namely, random forests, support vector machines and multiple kernel learning on 20 cancer types for two experiment types. Our MLP algorithm obtained better or comparable predictive performance against other baseline algorithms on most of the TCGA datasets. The results showed that our algorithm can extract meaningful and disease-specific key mechanisms of cancer prognosis.