Gene and Network Analysis of Common Variants Reveals Novel Associations in Multiple Complex Diseases

Abstract

Genome-wide association (GWA) studies typically lack power to detect genotypes significantly associated with complex diseases, where different causal mutations of small effect may be present across cases. A common, tractable approach for identifying genomic elements associated with complex traits is to evaluate combinations of variants in known pathways or gene sets with shared biological function. Such gene-set analyses require the computation of gene-level P-values or gene scores; these gene scores are also useful when generating hypotheses for experimental validation. However, commonly used methods for generating GWA gene scores are computationally inefficient, biased by gene length, imprecise, or have low true positive rate (TPR) at low false positive rates (FPR), leading to erroneous hypotheses for functional validation. Here we introduce a new method, PEGASUS, for analytically calculating gene scores. PEGASUS produces gene scores with as much as 10 orders of magnitude higher numerical precision than competing methods. In simulation, PEGASUS outperforms existing methods, achieving up to 30% higher TPR when the FPR is fixed at 1%. We use gene scores from PEGASUS as input to HotNet2 to identify networks of interacting genes associated with multiple complex diseases and traits; this is the first application of HotNet2 to common variation. In ulcerative colitis and waist–hip ratio, we discover networks that include genes previously associated with these phenotypes, as well as novel candidate genes. In contrast, existing methods fail to identify these networks. We also identify networks for attention-deficit/hyperactivity disorder, in which GWA studies have yet to identify any significant SNPs.

Novel Gene and Network Associations Found for Acute Lymphoblastic Leukemia Using Case–Control and Family-Based Studies in Multiethnic Populations

“…To identify genes associated with ALL in each of the two case-control datasets and the trio analysis dataset, we performed gene-level tests of association using PEGASUS (20). Briefly, individual SNP statistics are drawn from a chi-square distribution correlated by empirical LD, and the distribution of the sum of correlated chi-square statistics within a gene is the null distribution for gene-level statistics; this distribution is then numerically integrated to calculate a gene-level p -value with machine precision (20).…”

“…Briefly, individual SNP statistics are drawn from a chi-square distribution correlated by empirical LD, and the distribution of the sum of correlated chi-square statistics within a gene is the null distribution for gene-level statistics; this distribution is then numerically integrated to calculate a gene-level p -value with machine precision (20). We calculate gene-level p -values or “scores” for 19,000 genes using gene boundaries of 50,000 bp upstream and downstream of the genes to account for regulatory regions; gene start and end positions are downloaded from the UCSC Genome Browser.…”

“…HotNet2 uses directed heat diffusion along interaction networks where every gene, or a “node” in the network graph, has a “heat score” based on its gene score. Though originally developed for analyzing somatic mutation data from cancer datasets, HotNet2 has been used to uncover gene subnetworks significantly associated with common traits and diseases using p -values from GWA studies on common variants (20). We use negative log-transformed gene scores generated by PEGASUS as heat scores in HotNet2.…”

“…In the gene-level method, PEGASUS (20), association p -values are combined within genes while correcting for linkage disequilibrium (LD); this approach allows us to account for genetic heterogeneity -- when different causal mutations of small effect in the same gene or pathway may be present across cases -- and to test for genes and pathways significantly associated with ALL. Because several ALL susceptibility genes are known to play integral roles in lymphoid development, cell type differentiation and leukemogenesis, it is important to determine if these genes act in concert or separately in ALL patients.…”

“…We used the gene-level association method PEGASUS (20) to identify novel candidate genes associated with ALL. We then applied the gene-set enrichment method DAVID (17,18) to identify enriched functional categories in PEGASUS-identified gene-level associations with ALL.…”