Network analysis of GWAS data

Abstract

Genome-wide association studies (GWAS) identify genetic variants that distinguish a control population from a population with a specific trait. Two challenges in GWAS are: (1) identification of the causal variant within a longer haplotype that is associated with the trait; (2) identification of causal variants for polygenic traits that are caused by variants in multiple genes within a pathway. We review recent methods that use information in protein–protein and protein–DNA interaction networks to address these two challenges.

Introduction

Genome-wide association studies (GWAS) aim to identify genetic variants that distinguish a population of individuals, or cases, that have a particular phenotype/trait (typically a disease) from control individuals [1]. In its simplest form, analysis of a GWAS is a logistic regression where for each genotyped single-nucleotide polymorphism (SNP) the number of copies of the non-reference allele is regressed onto disease status for all individuals. The resulting P-value for each SNP is then corrected for multiple tests, and SNPs with alleles significantly enriched in controls are identified (Figure 1a).

Figure 1.

Two applications of network-based analyses of GWAS. (a) GWAS analysis computes the association between a SNP and case/control, reporting a P-value for each SNP. (b) Casual gene identification is the problem of identifying a single causal gene (circled in red) for the phenotype from a larger locus of candidate genes that is significantly associated with the phenotype. (c) Causal network identification is the problem of finding a group of interacting genes (e.g. a signaling pathway or protein complex) containing SNPs that distinguish cases and controls.

There are two major challenges in using GWAS to identify the genomic underpinnings of complex phenotypes (Figure 1). First, GWAS-identified SNPs are generally not located in the gene(s) underlying the phenotype of interest, but rather, are in linkage disequilibrium with causal genes or SNPs. Thus, one challenge is to identify causal genes within a GWAS-implicated locus (Figure 1b). One solution to this challenge is to use interaction networks to rank genes within a haplotype according to interactions with other genes known to be associated to the phenotype of interest or to similar phenotypes.

A second challenge is that GWAS-detected variants do not explain most of the genetic effects found in affected individuals – even for diseases known to have a strong genetic component, such as obesity and diabetes. This has been termed the “missing heritability problem”) [2–5]. An underexplored cause of missing heritability is genetic heterogeneity: the concept that different collections of causal variants are present in different patients. Genetic heterogeneity manifests itself on two levels. First, affected individuals may harbor distinct causal variants within a given causal gene. Second, causal variants may be distributed across different genes within a pathway (signaling, regulatory, metabolic) or protein complex [6]. This review focuses on the second type of genetic heterogeneity.

Genetic heterogeneity resulting from pathways and protein complexes complicates GWAS because for any specific causal gene, only a subset of the cases will contain a variant in that gene, while other cases will have causal variants in other genes in the pathway. This reduces the power of tests of association between single genes and the phenotype. Unraveling such genetic heterogeneity requires testing the association between the phenotype of interest and different combinations of genes containing putative causal variants. The goal is to identify sets of genes with the property that each affected individual contains a causal variant in at least one gene in the set. It is also possible to consider the case where an affected individual contains multiple causal variants in different genes in the set, but we will not consider this case here. The naive approach of exhaustively testing all combinations of variants is not computationally or statistically feasible. For example, one cannot exhaustively test all 10²⁰ combinations of 5 genes and retain statistical power without data from an astronomical number of individuals.

In this review, we describe recent work using interaction networks to address these two challenges in GWAS, focusing on three specific applications:

1. Causal gene identification. It has been observed that different causal genes for the same or similar phenotypes often interact, either directly or via common interaction partners. Network approaches use this observation to select putative causal gene(s) from haplotypes by finding genes that are close or related in a network to other known causal genes.

2. Causal gene identification for expression phenotypes. pt?>Gene expression is a phenotype of particular interest because it is readily measured from micro-arrays or RNA-Seq. Because gene expression is a molecular phenotype, network approaches are attractive as they may provide a mechanistic explanation for a causal variant.

3. Causal network identification. GWAS of genetically heterogeneous or polygenic diseases require testing groups of genes that are known to participate in the same biological process. Standard gene set enrichment or ranking statistics have been used to test known pathways in GWAS [6]. Interaction networks provide an alternative source of information that can be used profitably to identify combinations of causal variants without limiting analysis to known pathways.

In this review, we focus on the use of interaction networks in GWAS, and more specifically in common variant association studies (CVAS). However, we also briefly summarize some of the approaches used for the analogous causal network identification problem in cancer genome sequencing studies [7,8].

Network approaches

Interaction networks

Large-scale interaction networks incorporate the results of both molecular and high-throughput experiments to describe different biochemical relationships between genes and the proteins they encode. These networks take the form of a graph G = (V, E). The vertices V represent genes and their corresponding protein products. The edges E join pairs of vertices whose corresponding proteins exhibit a specific biochemical interaction (e.g. physical association, phosphorylation, etc.). In some cases, the edges may have a direction corresponding to the directionality of the biological interaction. Commonly used protein–protein interaction (PPI) networks include HPRD [9], BioGRID [10], STRING [11], iRefIndex [12], and Reactome [13], most of which combine literature-curated interactions and interactions derived from high-throughput experiments [14–18]. More recently, Multinet [19^•] also integrates protein–DNA interactions from ENCODE.

Causal gene identification

The most common use of interaction networks in GWAS analysis is to identify the causal gene inside a haplotype block (Figure 2 and Table 1a). While GWAS identify haplotype blocks associated with a particular disease or phenotype, they typically do not have the resolution to identify the causal gene within the associated block. A network approach to causal gene identification is motivated by the observation that the protein products of causal genes often directly interact with, or share many interacting partners with, the protein products of other causal genes for the disease. Thus, given prior knowledge of causal genes for a phenotype, one may identify new causal genes by finding the gene in the haplotype block that is closest on the network to the known causal genes.

Figure 2.

Schematic of methods for causal gene identification. (a) Candidate causal genes in a locus (or haplotype block) identified as significantly associated with a phenotype by a GWA study are mapped (blue circles) to a protein–protein interaction network. Each candidate gene is ranked in relation to a set of known causal genes (green squares; for simplicity, only one causal gene is shown) using a network distance measure. Different network distance functions that incorporate different features of network topology have been proposed including connectivity (e.g. direct interactions), network flow, random walks, and topological similarity (e.g. diffusion “profiles”). (b) Methods for identifying causal genes for expression phenotypes identify a causal gene from a locus of candidate genes (blue circles) that explain a differentially-expressed gene (red circle). Network methods find explanatory path(s) from the causal gene to the differentially expressed gene through an integrated network of protein–protein and protein–DNA interactions that provide a mechanistic explanation for the change in expression. In this example, candidate gene s is identified as upstream of differentially-expressed gene G₄ with explanatory path (blue) from s to G₄ terminates in a protein–DNA interaction.

Table 1.

Network analysis methods for GWAS

Algorithmic approach	Reference	Interactome	Genetic/phenotypic data
a. Causal gene identification
Direct neighbors	Oti et al. [20]	HPRD + high-throughput experiments	Causal genes
	CIPHER [21]	HPRD + OPHID + BIND + MINT	Causal genes + phenome
	Lee et al. [55]	HumanNet	GWAS SNPs
Network flow & random walks	GeneWanderer [23]	HPRD, BIND, BioGrid, IntAct, DIP, STRING	Causal genes
	PRINCE [24]	HPRD + high-throughput experiments (weighted)	Causal genesa + phenotype similarity scores
	MAXIF [29]	HPRD	Causal genes + phenome
	Zhu et al. [25]	HPRD	Causal genes + phenome
Topological similarity	AlignPI [27]	HPRD	Causal genes + phenome
	VAVIEN [26]	NCBI Entrez Gene (weighted)	Causal genes + phenotype similarity scores
b. Causal gene identification  for expression phenotypes
Topological properties	Kreimer and Pe’er [35]	HPRD	eSNPs
Network flow	Tu et al. [30]	PPI: yeast	eQTLs
		PDI: yeast
	ResponseNet [32]	PPI: yeast	eQTLs
		PDI: yeast (weighted)
	ResponseNet2.0 [33]	PPI: BioGRID + DIP + MINT + IntAct	eQTLs
		PDI: TRANSFAC (weighted)
Conductance	eQED [31•]	Yeast (weighted)	eQTLs
	Kim et al. [34]	PPI: MINT + IntAct + Reactome + HPRD + others PDI: TRED	eQTLs
c. Causal network identification
Seed and extend	PINBPA [36,42•]	iRefIndex filtered for high-confidence interactions	GWAS SNPs
	dmGWAS [41]	MINT + IntAct + DIP + BioGRID + HPRD + MIPS	GWAS SNPs
	NETBAG [39]	BIND + BioGRID + DIP + HPRD + InNetDB + IntAct + BiGG + MINT + MIPS	De novo CNVs
	NETBAG + [40•]	BIND + BioGRID + DIP + HPRD + InNetDB + IntAct + BiGG + MINT + MIPS	De novo CNVs + SNVs + GWAS-implicated loci
Exhaustive search of 2-step networks	NIMMI [44]	BioGRID	GWAS SNPs

^aGeneCards is the source of causal gene information for PRINCE. For all other methods, OMIM is the source of causal gene information.

Early methods used a simple definition of network distance, examining only nearest neighbors on a protein interaction network [20,21]. However, most biological interaction networks have a heavy-tailed degree distribution [22], meaning that most pairs of proteins are connected via short paths. This property makes nearest neighbors or shortest paths less desirable distance measures. The first method to utilize a more sophisticated measure of network distance that considers the overall topology of the network, GeneWanderer [23], ranks candidate genes based on the probability that a random walk starting from a known disease gene will finish at each candidate gene. Similar approaches measure network distance using information flow and network propagation [23–25].^a Two other methods select candidate causal genes based on their topological similarity to known causal genes [26,27] rather than their network distance.

Several of these methods also improve upon early approaches by incorporating phenotype similarity scores between diseases based on the overlap of their OMIM medical subject heading descriptions (described in [28]). Some methods incorporate phenotype similarity scores only for disease pairs including the disease for which causal genes are sought [24,26], while others integrate a “phenome” network in which phenotypes are nodes and weighted edges between all phenotype pairs represent their similarity [21,25,29]. Incorporating this information enables causal gene identification for diseases for which there are no previously known causal genes.

Causal gene identification for expression phenotypes

An important subproblem of causal gene identification arises when the phenotype of interest is gene expression; loci associated to a gene expression phenotype are sometimes referred to as expression quantitative trait loci (eQTL) or expression SNPs (eSNPs). Network approaches have been used to provide mechanistic explanations for observed correlations between a locus containing one or more source genes and a target gene that is differentially expressed between cases and controls (Table 1b). These methods find high-scoring paths in a combined protein–protein and protein–DNA interaction (PDI) network between one of the source genes and the target gene (Figure 2b). To explain the change in expression, the final edge in these paths is a protein–DNA interaction between a transcription factor that regulates the target gene. Three of the first such methods [30,31^•,32] analyzed eQTLs in yeast. The eQED algorithm of [31^•] used an electrical resistance model to find high-weight explanatory paths that connect SNPs to differentially expressed genes through known signaling and regulatory interactions. In comparison, ResponseNet [32] and ResponseNet2.0 [33] formulate the problem as a minimum-cost network flow, which is mathematically related to electrical resistance. Kim et al. [34] further extended these ideas, applying them to human cancer data and adding additional steps to identify causal genes from multiple explanatory paths. More recently, Kriemer [35] analyzed eSNPs identified in human whole-genome and RNA-Seq data, and found that source and target genes are generally closer on the PPI network. However, in contrast to the work above, they did not use protein–DNA interactions to find explanatory paths for these associations.

Causal network identification

A third use of interaction networks in GWAS analysis is to identify causal networks, or sets of interacting genes containing causal variants. This approach complements popular pathway-based tests that restrict attention to groups of variants in known pathways or gene sets using enrichment statistics [6,36,37^•]. Network approaches address three limitations of gene set analysis. First, gene sets do not model the topology and type of interactions between genes, and instead treat all genes in the set as equivalent. Second, gene set methods perform a separate statistical test on each gene set and do not consider the interconnection of pathways in larger signaling and regulatory networks. Third, by restricting attention to known pathways, gene set methods are unable to discover novel groups of interacting genes that are associated to the phenotype.

Several algorithms have been introduced to find causal networks in protein–protein interaction networks (Figure 3a and Table 1c) [36,38,39,40^•,41,42^•]. Authors [36,42^•] use the jActiveModules plug-in [43] in Cytoscape to analyze multiple sclerosis GWAS data on the iRefIndex protein–protein interaction network [17]. jActiveModules provides a general approach to find high-scoring subnetworks in a vertex-weighted network (Figure 3b). dmGWAS is a similar approach [41]. The NETBAG [39] and NETBAG + algorithms [40^•] – used to identify subnetworks affected by rare and de novo variants in autism and schizophrenia, respectively – are also related but analyze an edge-weighted interaction network. All of these methods use a greedy heuristic (“seed and extend”) to find high-scoring subnetworks by iteratively adding to a subnetwork those genes that increase the subnetwork’s score (Figure 3b). These approaches compute the statistical significance of the resulting subnetworks by comparing to an empirical distribution of subnetwork scores.

Figure 3.

Schematic of methods for causal network identification and examples of two algorithms. (a) Proteins in the protein–protein interaction network are scored using the association P-values within or near their corresponding gene. In this example, nodes are colored using a blue-to-red gradient where blue represents low scores and red represents high scores. Proteins without scores (i.e. those that were not tested in the GWA study or had no significant associations) are colored gray but remain in the network for analysis due to their effect on the network’s topology. High-scoring subnetworks are then reported, taking into account both the protein scores and the network topology. (b) jActiveModules, NETBAG, and NETBAG + all use a greedy heuristic (seed and extend) to identify causal networks by iteratively adding to a subnetwork genes that increase the subnetwork score. jActiveModules uses a vertex-weighted graph where each vertex has an associated Z-score, and the score of a subnetwork with k nodes is the normalized sum ${\sum }_i{Z}_i∕\sqrt{k}$ of Z-scores. In the original application of jActiveModules, the Z-score of a gene indicated its differential expression in microarray experiments. For the application to GWAS, [36,42^•] transform gene-level P-values (from VEGAS [46^•]) of association into Z-scores. NETBAG algorithms [39,40^•] analyze a weighted graph with edge weights determined by naïve Bayes integration of protein interaction and protein complex databases, protein sequence alignment, and co-evolution. In the vertex-weighted graph shown, G₁ is the seed gene, and genes G₄, G₅, G₂, and G₆ are added to the subnetwork in that order (as indicated with labels on the edges) G₃ is not added because it has a low score. (c) HotNet uses heat diffusion in order to identify causal networks. Heat is assigned to each gene in proportion to its score and diffuses over the edges of the network. The heat diffusion process takes into account the topology of the network so that genes with high-degree pass proportionally less heat to their neighbors than genes with low degree. In the example shown, G₄ and G₃ are initially cold (indicated by light blue), while G₁ and G₂ are “hot” (indicated by red and orange, respectively). After heat diffuses along the edges, G₁, G₂, and G₄ have the same heat, while G₃ is colder than G₄ because it is not directly connected to G₁. The remaining nodes G₅ and G₆ are initially cold and remain cold because they are only connected to the high-degree G₁. A hot subnetwork of genes G₁, G₂, G₃, and G₄ is identified.

An additional approach is the Network Interface Miner for Multigenic Interactions (NIMMI) [44]. NIMMI employs a modified version of the PageRank algorithm for webpage ranking [45] to compute a weight for each gene that represents its network centrality. These weights are combined with gene-wise P-values from VEGAS [46^•], and an exhaustive search is performed of all subnetworks consisting of paths of length 2 from a starting node.

In comparison to the number of methods for causal gene identification, there remain relatively few methods for causal network identification. However, an analogous problem occurs in cancer genome sequencing studies, where the challenge is to identify signaling/regulatory/metabolic networks harboring more somatic aberrations than expected by chance [7,8]. One algorithm introduced for this task, NetBox [47], decomposes a network into modules of mutated genes that are either directly connected or connected through single linker genes. Another algorithm, HotNet [48], uses a heat diffusion model to identify significantly mutated subnetworks as “hotspots” on the network (Figure 3c). Heat is assigned to each node in proportion to its mutation frequency, and this heat then diffuses over the edges of the graph, either for a fixed time [49] or until equilibrium [48]. Hot subnetworks are found by removing cold edges and the statistical significance of the number and size of the resulting hot subnetworks is computed. Thus, HotNet simultaneously considers both the score assigned to each gene and the global topology of the network, in contrast to most of the methods above that use these two features sequentially. Despite the generality of these two algorithms, neither has yet been used to analyze GWAS data. We discuss prospects for adapting these methods for GWAS analysis in the next section.

Challenges and future prospects

A number of challenges remain in network analysis of GWAS. First, network methods are limited by the coverage and quality of protein–protein and protein–DNA interaction networks. High-quality experimental interaction data are laborious to obtain. Consequently, existing network databases have many missing interactions, and these reduce the sensitivity of network analyses. High-throughput interaction data, combined with additional experimental validation, will be crucial to increase sensitivity. Conversely, interaction databases also contain false positive interactions. Some of these are a result of incorrect predictions, errors in data curation, or experimental noise. Others result from the fact that most interaction networks are a superposition of interactions measured in different cell types and conditions, only a subset of which may be active in the tissue of the disease. Authors of [50,51] demonstrated that tissue-specific protein–protein interaction networks can improve disease-gene prioritization results.

Second, the dramatic decline in the cost of DNA sequencing is enabling whole-exome and whole-genome sequencing of cases and controls. Sequencing allows the analysis of de novo variants and rare variants in both coding and non-coding regions. A promising exampleof this type of analysis is demonstrated by Gulsuner et al. [52^••], who identified causal subnetworks of interaction networks that contain significant numbers of de novo variants in schizophrenia patients. However, the challenge of extending causal network and causal gene identification approaches to rare variants requires additional methodological advances. For example, since causal rare variants may be randomly associated with different common haplotypes in sampled individuals, most rare variant association study (RVAS) analyses require sensible methods to pool variants across a gene or locus [53]. These approaches help address the problem of genetic heterogeneity resulting from different causal variants within a specific causal gene, but leave open the issue of rare causal variants across genes in a pathway/complex. A combination of pooled rare variants within a locus and network approaches across a locus is a promising direction.

In addition to a role for network approaches in CVAS, RVAS and de novo variant studies, network analyses have proven useful in the analysis of somatic mutations in cancer genomes. Cancer genome sequencing studies face an analogous problem of genetic heterogeneity where causal somatic mutations, or driver mutations, are distributed across multiple genes in a pathway [7,8]. As noted above, several network methods have been introduced for this problem [47–49]. While some of these methods may prove useful for germline variants, there are notable differences in the analyses of somatic vs. germline variants. First, somatic mutations, as well as de novo germline mutations, arise independently in each individual, and thus can be analyzed without considering ancestry and population structure. In contrast, analyses of common and/or rare variants require additional techniques to control for spurious associations with ancestry. Second, analysis of somatic mutations in tumors face issues such as intratumor heterogeneity that do not have parallels in germline studies. Despite these differences, both types of analyses can benefit from greater exchange of methodology.

Looking outside genes, network analysis of non-coding SNPs requires additional information about regulatory interactions, non-coding RNAs, among others. The ENCODE project [49] is an important first step in the generation of such information, but more data are needed. Network analysis will play an increasingly important role in prioritizing candidate causal variants for further experimental validation. Ultimately, the combination of computational and experimental approaches will yield mechanistic insights into the process by which a genetic variant, or a combination of variants, affect a complex phenotype.