Integrative clustering approaches for cancer subtype discovery

Cancer is a heterogeneous disease and identification of cancer subtypes is critical for personalized treatment and drug development. Recently, cancer genome projects have produced multiple types of high-throughput data for thousands of cancer patients. Exploiting the complementary information between different data types can improve finding subtypes. In the first part of this thesis, we apply multi-view kernel k-means to integrate multiple genomic datasets (i.e.,gene expression, DNA methylation and miRNA expression) on two cancer datasets. We show that combining kernels (i.e., that correspond to different views) with learned weights give better clusters compared to combining the kernels uniformly or using each data set independently. We also demonstrate an improved performance compared to existing models that integrate the same data types. In terms of biological significance, Kaplan-Meier analysis shows that our discovered clusters have distinct survival profiles with statistically significant log-rank test p-values. In the second part, we target the high dimensionality problem of genomic datasets by extending a single-view sparse k-means framework to multi-view setting. This extension allows us to perform integrative clustering and feature selection simultaneously by learning both feature weights and view weights. We confirm that performing feature selection improves the clusters for the majority of the datasets. Altogether, our results indicate that integration of multiple genomic characterizations and the application of feature selection enable the discovery of subtypes that improve over current patient stratifications.