A powerful and robust co-expression network analysis algorithm

Networks provide classic approaches to model complex systems, hence their wide application to investigate biological systems. Many types of biological networks are available; we have focused on co-expression networks. Co-expression networks have become standard tools in the systems biology toolbox to examine possible functional relationships in whole-transcriptome and -genome studies. These networks provide simple yet effective models of putative gene or protein interactions and potential functional groupings. Current methods often use parametric metrics, a misleading default option, as most high-throughput expression data are not normally distributed. To provide statistically sound network construction, our algorithm includes parametric and non-parametric similarity measures to allow for various types of data distributions. Biological networks are known to have approximate scale-free and small-world structure, thus we believe it is imperative that any co-expression network models follow these two properties, a feature often not provided or supported by current available methods. By allowing the network construction to be governed by these properties, our networks are guaranteed to be both scale-free and small-world. Our algorithm is implemented in R, with minimal effort required from the user. Input requirements consist only of a text file of quantified expression measures, and the choice of similarity metric. The output is a simple text file indicating connections between all nodes in the expression data file. This file can then be used for downstream analysis, such as the identification of interesting network structures. These structures include, but are not limited to, paths, cliques, modules, and clusters, which provide biological insights and possibly aid in hypothesizing novel functional properties of genes within the network. For example, paths can indicate actual metabolic pathways and cliques may illuminate functional properties of novel genes within gene groupings. One of our goals is to identify cliques within the co-expression network and use their mathematical properties to identify stronger functional gene groupings than typical clustering routines. Additionally, we hope that cliques will enable us to reduce the dimensionality of the network by merging genes (nodes) belonging to the same cliques into 'bigger' nodes. This will benefit in the visualization of the network.

We believe an appropriate statistical sound similarity measure in combination with the known network properties will best enable the construction of mathematically reliable and biologically meaningful and relevant networks. Our main goal therefore is to generate more biologically meaningful network models, which then can be examined for connectivity structures to identify gene modules, functional groups, and regulatory/influential patterns, which may be missed by current methods. Finally, we propose to apply these approaches to whole-genome data and whole-transcriptome experiments unlike other algorithms utilizing only a pre-selected gene subset.