Visualizing Personalized Cancer Risk Prediction

Cancer, as we now know, is a genetic disease. This knowledge entails that to enable evidence-based personalized diagnosis for any patient, the location of where the tumor occurs (e.g. brain, liver, etc.) is less relevant than the underlying genetic signature that the cancer cells express (i.e. whether genes are silenced, amplified or mutated). With the advent of high-throughput gene sequencing technologies, it has been possible to sequence the entire human genome of disease-diagnosed or normal patients. As a consequence, large repositories of genomic datasets are now available for data analysis, and identifying the underlying patterns could aid in the diagnosis, prognosis and treatment on a personalized basis. The Cancer Genome Atlas (TCGA) publishes data pertaining to the molecular information (Exon (mRNA) expression, DNA Methylation, Single Nucleotide Polymorphisms (SNP), Copy Number Variations (CNV) etc.) and clinical attributes of around 9000 patients tested across different cancer typologies. Data from TCGA is of high value for oncologists as it enables matching the genomic evidence found in their own patients with those enrolled in the TCGA project. TCGA data has been widely used for hypothesis-driven translational research as all of its data results are from direct experimental evidence. Previous research has been carried out for the discovery of new tumor bio-markers (genes or single nucleotides) using unsupervised learning algorithms, and classification of patient samples using supervised learning methods. However most of these analyses use TCGA Exon Expression or SNP datasets. It has been shown recently that DNA methylation signatures are robust bio-markers, vastly more stable than mRNA or proteins, and hence will extend our ability to classify cancer and predict outcome beyond what is currently possible. This could lead the development of new approaches for diagnosis and prognosis of different kinds of cancer [1]. However, class prediction based on these patterns is an under-determined problem, due to the extreme high dimensionality of the data compared to the usually small number of available samples. Hence, a reduction of the data dimensionality, through a combination of several feature selection methods, is a necessary pre-processing step for classification performance.