Guilt by rewiring: gene prioritization through network rewiring in genome wide association studies

Abstract

Although Genome Wide Association Studies (GWAS) have identified many susceptibility loci for common diseases, they only explain a small portion of heritability. It is challenging to identify the remaining disease loci because their association signals are likely weak and difficult to identify among millions of candidates. One potentially useful direction to increase statistical power is to incorporate functional genomics information, especially gene expression networks, to prioritize GWAS signals. Most current methods utilizing network information to prioritize disease genes are based on the 'guilt by association' principle, in which networks are treated as static, and disease-associated genes are assumed to locate closer with each other than random pairs in the network. In contrast, we propose a novel 'guilt by rewiring' principle. Studying the dynamics of gene networks between controls and patients, this principle assumes that disease genes more likely undergo rewiring in patients, whereas most of the network remains unaffected in disease condition. To demonstrate this principle, we consider the changes of co-expression networks in Crohn's disease patients and controls, and how network dynamics reveals information on disease associations. Our results demonstrate that network rewiring is abundant in the immune system, and disease-associated genes are more likely to be rewired in patients. To integrate this network rewiring feature and GWAS signals, we propose to use the Markov random field framework to integrate network information to prioritize genes. Applications in Crohn's disease and Parkinson's disease show that this framework leads to more replicable results, and implicates potentially disease-associated pathways.

Figure 1.

Network rewiring is abundant in immune system (defined in the Reactome pathway). The y-axis is the fold change of density in the subnetwork of the immune system to the total network. Network density is the number of edges in the observed network over the number of all possible edges.

Figure 2.

Comparisons of edge frequency in static network and rewiring network. (+1, +1) denotes an edge that connects two associated genes, while (+1, −1) denotes an edge that connects one associated gene and one non-associated gene. The y-axis is the fold change between observed number of edges and that of random expectation. Both the static and rewiring network were dichotomized at a threshold so that the corresponding network density is 0.01.

Figure 3.

Replication rates between independent cohorts in Crohn's disease study. A gene is called replicable if its association P-value in the replication cohort is <0.05. (A). Black circle: prioritized genes. Grey circle: non-prioritized genes. Black diamond: baseline replication rates for the top k genes. Grey cross: prioritized genes with microarray data set GSE8397 (a microarray data set of Parkinson's disease). The numbers on the top of x-axis are the association P-value of the kth genes in the discovery cohort. (B). Empirical P-values of the replication rate of the prioritized gene set, derived from permutations. The dotted line indicates significance level at 0.05. The y-axis is drawn at the log₁₀ scale but labelled with original P-values for ease of reading.

Figure 4.

Replication rates between independent cohorts in Parkinson's disease study. A gene is called replicable if its association P-value in the replication cohort is <0.05. (A) Black circles: prioritized genes. Grey circles: non-prioritized genes. Black diamonds: baseline replication rates for the top k genes. Grey crosses: prioritized genes with microarray data set GSE20881 (a microarray data set of Crohn's disease). The numbers on the top of x-axis are the association P-value of the kth genes in the discovery cohort. (B) Empirical P-values of the replication rate of the prioritized gene set derived from permutations. The dotted line indicates significance level at 0.05. The y-axis is drawn at the log₁₀ scale but labelled with original P-values for ease of reading.