A new integrative data mining framework for analyzing the cancer genome atlas data

Besides accuracy and efficiency, understandability is another key issue of predictive modeling in real-world applications, especially in biomedical and healthcare data analysis. We develop a new integrative framework to enhance the interpretability of data by sparsity-based learning. We proposed several novel sparsity-based learning models, emphasizing different understandable properties of data, such as explicit sparsity, low redundancy, and low rank, and apply to The Cancer Genome Atlas (TCGA) data analysis. Results indicate that the proposed methods provide more insights from TCGA data while maintaining stable and competitive performances in predictive modeling. To further enhance the interpretability of biological processes and disease mechanisms, we also develop a novel visualization tool by considering heterogeneous relationships among genomics elements. By applying the novel learning models and the visualization tools, pathways of several important cancer diseases are revisited and a series of novel potential bio-markers are discovered which improves our ability to diagnosis, treat and prevent cancer.