Extracting local information for identifying differentially expressed pathways

This paper proposes to extract local information in a gene set for identifying differentially expressed (DE) gene pathways. DE pathways are more meaningful than a single DE gene to understanding a biological process, and identifying DE pathways has drawn more and more attentions recently. Current methods are mainly based on the identification of single DE genes, and do not concern correlations between genes. We propose to extract local correlations in a pathway of interest by randomly sampling multiple gene subsets from it and using a logistic regression model to measure how the local correlation pattern in each subset predicts phenotypic labels. The differential expression significance of the pathway is finally assessed by combining the p-values of the subsets predicting phenotypic labels to a combinative one. The proposed method referred to as locLR is evaluated on three simulation data sets and a real-world data, and is shown to be more powerful for identifying DE pathways than the previous methods.