Network-based Wavelet Smoothing for Analysis of Genomic Data

Abstract

Biological networks are a common way of describing information on relationships between genes that are accumulated from many years of biomedical research, and they are thus potentially valuable when incorporated as prior knowledge to guide biomarker discovery in genomic data analysis. In this study, we focus on network-based regularization methods through a predictive framework with linear models, and propose to use a class of methods based on wavelet smoothing over undirected graphs that directly detect subnetworks composing of collaboratively functional gene modules. We perform breast cancer survival analysis using a large gene expression dataset and a protein-protein interaction network obtained from a public database, and demonstrate that the proposed methods are able to improve gene selection in terms of stability, connectivity and interpretability while achieving competitive performance of survival risk prediction. Our results also serve a comparative study benchmarking several network-free and network-based regularization methods for gene selection related to breast cancer survival.