Exploratory analysis of local gene groups in breast cancer guided by biological networks

The path to personalized medicine requires the stratification of patients based on their genetic, molecular, and other characteristics to achieve the individualized treatment of complex diseases such as the breast cancer. The identification of single “biomarkers” as the driving forces for the appearance of cancer has therefore been widely pursued in the last fifteen years but with no robust results across different studies. The use of existing biological knowledge such as the gene interaction networks and regulatory pathways can be of great help, since it has been argued that cancer is caused by the deregulation of multiple biological processes in the cell. In this study we explore the usage of such biological knowledge for the tuning and adaptation of the breast cancer classification tasks both in a supervised (classifying unknown samples according to a predetermined taxonomy) and unsupervised setting (clustering of new data towards identifying new categories). The proposed methodology starts from an initial list of “seed” genes and proceeds to the expansion of their “neighborhoods” according to the topology of a given biological network. The expansion process operates in a supervised manner for the construction of the first level in a two level classification scheme. The first level base classifiers are built using the network structure and a “random walk” search strategy for the selection of the genes used in these classifiers. At the second level, a meta-classifier is trained to combine in the best possible way the results of the base classifiers. The proposed approach therefore aims to strengthen the predictive ability of the initial list of genes and provide more robust generalization guarantees. Proceeding to the unsupervised setting, the extracted gene neighborhoods around the initial “seeds” are considered as modules of highly interacting genes within the same group but of strong independence across groups. This consideration allows the introduction of a sparse Gaussian mixture model for the assignment of breast cancer samples into a set of unknown clusters. Our methodology is explained in full detail and promising results in Breast Cancer related scenarios are obtained.