Pathway expression analysis

Abstract

This paper introduces a pathway expression framework as an approach for constructing derived biomarkers. The pathway expression framework incorporates the biological connections of genes leading to a biologically relevant model. Using this framework, we distinguish between shedding subjects post-infection and all subjects pre-infection in human blood transcriptomic samples challenged with various respiratory viruses: H1N1, H3N2, HRV (Human Rhinoviruses), and RSV (Respiratory Syncytial Virus). Additionally, pathway expression data is used for selecting discriminatory pathways from these experiments. The classification results and selected pathways are benchmarked against standard gene expression based classification and pathway ranking methodologies. We find that using the pathway expression data along with selected pathways, which have minimal overlap with high ranking pathways found by traditional methods, improves classification rates across experiments.

Module representatives for refining gene co-expression modules

“…Pathway expression is a methodology for dimensionality reduction of gene expression data which translates the gene expression feature space into a pathway expression feature space [22]. Linear pathway expression methods generate a pathway expression value by averaging gene expression levels for the genes in the pathway.…”

“…In this paper, we will use correlation between genes expression levels to generate edges for our module gene network and PageRank centrality [27] for our centrality measure. We choose this because PageRank centrality with a correlation network produced a higher inter-quartile range of balanced success rate (BSR) than degree centrality in the initial pathway expression analysis paper [22]. With these parameters, the computational complexity of module expression for one module is the same as the computational complexity of PageRank.…”

“…The focus of this paper is on developing new methods for the representation of modules, analyzing their properties, and implementing them in a k-means-type Linde-Buzo-Gray (LBG) clustering algorithm for gene co-expression module refinement. We present a methodology which seeks to improve upon the WGCNA method for module representation by introducing eigengene subspaces, flag mean [20], flag median [21], and module expression vectors [22]. We seek to develop methods which improve biological significance of the resulting modules, at the possible expense of Machine Learning classification rates.…”

Using machine learning to determine the time of exposure to infection by a respiratory pathogen

“…In particular, one may endeavor to identify all of the biomarkers that are discriminatory for a given host immune response using Iterative Feature Removal , a repeated application of the sparse removal of top biomarkers until none remain 6 . This approach lends itself well to construction of predictive models, but also provides a potentially comprehensive picture that might lead to biological insights, e.g., in Lyme disease 7 , and influenza 8 ; or more generally, to data driven biological pathway analysis 9 . Of specific interest to us in this investigation is the fact that the discriminatory biomarkers associated with respiratory infections appear to be highly time-dependent 8 .…”