Cancer Subtype Detection Using Multi View Structural Kernels on Metabolic Pathways
Cancer subtype detection using genomic data tries to separate patients into several groups of different properties. This property can be lethality of the subtype or drug responsiveness. Studies that focus on subtype detection uses several different data sets such as metabolic pathways, RNASeq data, somatic mutations, and clinical data. Metabolic pathways gives us the relationship between genes and these gene nodes can be used to represent their role in the graph as a whole. We can use the structural role of the genes to generate graph kernels and map other gene related information to cluster our patients depending on the cluster properties we want to focus on. Node2Vec is a method to convert nodes to feature vectors using the structural variation in the graphs and allows us to simplify neighbor relationships of the node into a vector. We use these vectors to map genes and use a wider range of clustering algorithms.
Introducing data sets of different types and properties provides a better explanation for the underlying clusters of cancer patients but also introduces the problem of merging these data sets together. In order to merge these data we propose a method to use the Pathway Kernels to convert multiple types of genomic data and Multi-View Kernel clustering methods to cluster our patients into groups of significance. This way we can both use the structural information of the genes and the additional information from other data sets without bias from different properties and aim to find cancer subtype that provide more information