Learning Kernels from genetic profiles to discriminate tumor subtypes
Abstract
Several biological data types are getting easier to access. Genomic, Transcriptomic, Proteomic and Metabolomic data are some of the cases among others. Phenotype data also is included, since we can link any biological layer of information mentioned before with a phenotype, like the diagnosis of a disease. Feature selection methods help to determine the main genes and mutations responsible of characterize different Breast Cancer subtype tumors and to perform a better discriminant analysis. Tumor subtype classification has been performed using somatic point mutations [3] and support vector machines classifiers with standard and well-known kernels such us Gaussian. The methods commonly used to select features (genes) are Filter and Wrappers [8]. Our proposal is focused in learning kernels from human genomic data from Breast Cancer patients [4], in order to classify them among different tumor subtypes. Building custom Kernels provides an alternative to improve results. By combining different Kernel functions learned from the training samples we aim to improve the Kernel Target Alignment (KTA) score [1] and thus use the optimized kernel to perform discriminant analysis through support vector classification. Kernel Target Alignment measures the degree of agreement between a reproducing kernel function and a learning task. The higher the KTA, the better the performance of the support vector classification among different classes [2], thus a better discriminant analysis between tumor subtypes for the benefit of a better therapy. Kernel Target Alignment of the kernel 'K' and target label 'y' with respect to the sample 'S' of size 'm' is expressed as: (1) AGRANDA, Simposio Argentino de Grandes Datos 47JAIIO-AGRANDA