Cancerous Tissue Classification Using Microarray Gene Expression

In this project, we apply machine learning techniques to perform tumor vs. normal tissue classification using gene expression microarray data, which was proven to be useful for early-stage cancer diagnosis and cancer subtype identification. We compare the results of both supervised learning (k-nearest-neighbors, SVMs, boosting) and unsupervised learning (k-means clustering, hierarchical clustering) routines on three datasets: GSE3 (renal clear cell carcinoma), GSE8054 (pancreatic cancer) and SRBCT (round blue-cell tumors). In order to eliminate the non-informative genes from the data sets, we apply feature selection using the t-test, differential expression test, and the pairwise correlation coefficient between the class labels and each gene, which boosted the classification accuracy for most of the methods that we carried out. We present the misclassification error rate for hold-out (0.3), 5-fold, and leave-one-out cross validation. In agreement with prior work, we found machine learning techniques to be very efficient with this classification task. For binary classification (cancer vs. normal) the highest accuracy (close to 95% for GSE3 and more than 99% for GSE8054) was achieved with AdaBoost and a linear kernel SVM. For multi-class classification (SRBCT tumor subtypes) we achieve an accuracy of 100% with a linear kernel SVM without feature selection and 98% after reducing the feature dimension by 4 using the correlation coefficient feature selection technique.