Integrative Network-Based Machine Learning Approaches for Cancer Bioinformatics and Bio-Molecular Network Reconstruction

点击翻译按钮获取中文摘要

Network-based classifiers (NCs) that combine primary tumor data (e.g., gene expression data) with secondary data provided in the form of network (e.g., protein interaction networks or other bio-molecular networks) have been proposed for sample class prediction (e.g., cancer outcome prediction). The network data is used to identify informative groups of interacting genes, called network biomarkers (NBs), which can best separate the classes. Unfortunately, current NC methods proposed in cancer bioinformatics have produced very limited progress in terms of robust classification performance and stability of selected NBs. Current methods used for combining the primary tumor data with the network data do not well capture and summarize the biological knowledge contained within the data. The high dimensionality of the integrated data as well as the decoupling of the training of current NCs from the selection of genes hampers the stability of the NB identification. The performances of current NC methods are significantly hindered by the high level of noise and sparseness of current protein interaction networks. Network incompleteness is an important limitation in current NCs. My goal is to improve existing and devise new NC approaches which alleviate these limitations; by including additional biological data (tertiary data) useful for tumor classification, and then devising appropriate integration and learning methods which can best capture the biological information contained in the data. For a given cancer disease, finding accurate and robust predictive NBs will provide a better characterization of its subtypes, its outcomes, or its stages. The NBs will help physicians diagnose cancer more accurately, or suggest better treatment, and will also serve as potential drug targets in the future. The NBs, which describe the functional dependency between genes, can be monitored over time in the development of the disease, and hence, provide better strategies for cancer care and new strategies for the early detection of the disease. Current biomolecular networks such as Protein-Protein Interaction (PPI) networks are incomplete, contain many false-positive interactions and even many more false-negative interactions, and are very sparse with skewed degree distributions. This reduces the performance of many network-based prediction methods. I plan to propose network-based prediction methods which classify node pairs as interacting or not, in order to improve the quality of given networks. My approach will be based on the idea that "two nodes interact if they are closer to each other"; i.e., to propose and integrate different node similarity measures within learning frameworks. The proposed methods to improve the quality of networks are novel and the reconstructed PPI networks can be used in any network-based prediction problems (e.g., protein function prediction).