ABSTRACT

 

Motivation:  It has been long well known that genes do not act alone; rather groups of genes act in consort during a biological process. Consequently the expression levels of genes are dependent of each other. Experimental techniques to detect such interacting pairs of genes are in place for quite some time. With the advent of microarray technology, newer computational techniques to detect such interaction or association between gene expressions are being proposed which lead to an association network. While most microarray analyses look for genes that are differentially expressed, it is of potentially greater significance to identify how entire association network structures change between two or more biological settings, say normal versus diseased cell types.

 

Results:  We provide a recipe for conducting a differential analysis of networks constructed from microarray data under two experimental settings. At the core of our approach lies a connectivity score that represents the strength of genetic association or interaction between two genes. In this paper we use a score that is based on a partial least squares analysis of expression data that was introduced in our previous work. In this paper we use these scores to propose formal statistical tests for each of following queries: (i) whether the connectivity of a particular set of “interesting genes” has changed between the two networks, (ii) whether the connectivity of a given single gene has changed between the two networks and (iii)  whether the overall modular structures of the two networks is different. We carried out our method on two types of simulated data: Gaussian network and networks based on differential equations. We show that, for appropriate choices of the tuning parameters, our method works well on simulated data. We also analyze a real data set involving normal versus heavy mice and identify an interesting set of genes that may play key roles in obesity.