Eureka! Institutional Repository

The purpose of this thesis is to identify a small number of genes using Machine Learning methods, which contain sufficient information for the construction of a classifier capable of generalization. The dataset we examined concerns liver cancer and is from the TCGA database (The Cancer Genome Atlas). The target variables were the grade (Grades) and the stage of liver cancer (Stages). Gene expression data are transformed by RNA sequencing (RNA-seq). Initially, using the dimensional reduction methods PCA, Fisher Score, Mutual Information, Kolmogorov Smirnov 2 Samples, ReliefF, and mRMR MIQ, we evaluated the genes and ranked them based on their significance. Then, we chose the first 500 significant genes of each feature selection method and using the first 10 significant genes, we performed a Grid Search using the SVM Linear and SVM RBF classifiers to find the parameters that each classifier achieves the best mean accuracy value (20 Fold Cross Validation). We then repeated the above procedure by adding the next 10 significant genes each time up to 500. This process was also followed to examine the union and intersection subsets between all combinations of feature selection methods. The results of the experiments showed that the mRMR MIQ criterion is better than the other validation criteria we examined. With respect to both target variables, the mRMR MIQ criterion was able to identify the corresponding smaller number of significant genes that contain information capable of constructing an SVM RBF classifier that has the best generalization capability of all other subsets of important genes examined