Proteins Encoded in Genomic Regions Associated with Immune-Mediated Disease Physically Interact and Suggest Underlying Biology

Genome-wide association studies (GWAS) have defined over 150 genomic regions unequivocally containing variation predisposing to immune-mediated disease. Inferring disease biology from these observations, however, hinges on our ability to discover the molecular processes being perturbed by these risk variants. It has previously been observed that different genes harboring causal mutations for the same Mendelian disease often physically interact. We sought to evaluate the degree to which this is true of genes within strongly associated loci in complex disease. Using sets of loci defined in rheumatoid arthritis (RA) and Crohn's disease (CD) GWAS, we build protein–protein interaction (PPI) networks for genes within associated loci and find abundant physical interactions between protein products of associated genes. We apply multiple permutation approaches to show that these networks are more densely connected than chance expectation. To confirm biological relevance, we show that the components of the networks tend to be expressed in similar tissues relevant to the phenotypes in question, suggesting the network indicates common underlying processes perturbed by risk loci. Furthermore, we show that the RA and CD networks have predictive power by demonstrating that proteins in these networks, not encoded in the confirmed list of disease associated loci, are significantly enriched for association to the phenotypes in question in extended GWAS analysis. Finally, we test our method in 3 non-immune traits to assess its applicability to complex traits in general. We find that genes in loci associated to height and lipid levels assemble into significantly connected networks but did not detect excess connectivity among Type 2 Diabetes (T2D) loci beyond chance. Taken together, our results constitute evidence that, for many of the complex diseases studied here, common genetic associations implicate regions encoding proteins that physically interact in a preferential manner, in line with observations in Mendelian disease.

In identifying such processes, there are difficulties in both (i) identifying the specific genes at (and how they are molecularly impacted by) each association and (ii) inferring disease-causing mechanisms from the set of identified genes. Linkage disequilibrium blocks containing disease-associated SNPs can be hundreds of kilobases in size, and some contain tens of genes to consider. Genes are often informally implicated in pathogenesis by their proximity to the most associated marker, their biological plausibility, or simply their being the only protein-coding gene in the region. In reality, however, it is only a very small subset of confirmed GWAS associations for which specific functional variants have been proven experimentally.
More systematic approaches have been applied to connect genes to a common process with the use of independent data, such as Gene Set Enrichment Analysis (GSEA) and Gene Relationships Across Implicated Loci (GRAIL) [1,[19][20][21]. Both approaches identify connections between genes based on descriptive categories that outline the theorized underlying pathogenesis. However, these concepts are often general, so that specific hypotheses and molecular pathways can be difficult to define and are somewhat limited to established knowledge bases.
Observations of interactions between the products of proteincoding genes offer the most direct route to identifying pathogenic processes. It has been shown in a number of Mendelian diseases that genes causal of a particular phenotype tend to physically interact [22][23][24][25][26]. This has been confirmed in the model organism C. elegans, where RNAi knock-down of germline genes correlated highly with their products interacting in yeast-two hybrid experiments [26]. A classic example of a human Mendelian disease that recapitulates this model is Fanconi Anemia (FA), an autosomal recessive disorder linked to at least 13 loci, at least 8 of which function in a DNA repair complex [22]. Protein-protein interaction (PPI) data has also been used to formulate hypotheses about co-expressed genes as well as cancer genes [27,28]. We note that previous attempts to use PPI data to prioritize candidate genes in Mendelian disorders have been successful as was the case with the published tool Prioritizer [29]. We therefore set out to test such an approach in complex disease.
Investigators have rapidly populated databases of such proteinprotein interactions over the past decade. The reported interactions in PPI databases stem from both small, directed investigations and high-throughput experiments, primarily yeast twohybrid screens and affinity purification followed by mass spectrometry [30]. These data are inherently noisy: beyond technical false positives and negatives, experiments in vitro may report interactions that do not occur in vivo simply because the proteins involved never overlap spatially or temporally. To mitigate the noisiness of PPI databases, we extract networks from ''InWeb'', assembled in 2007 by Lage et al [24,31]. InWeb is a database of 169,810 high-confidence pair-wise interactions involving 12,793 proteins (human proteins and their orthologs). Lage et al. define high-confidence interactions as those seen in multiple independent experiments and reported more often in lower-throughput experiments [24]. To further restrict the data to biologically plausible interactions, we overlay mRNA expression information to confirm co-expression of binding partners; this correlates with co-localization, similar phenotype and participation in a protein complex [31,32].
Assessing the significance of networks built from PPI data is challenging for two reasons: first, overall connectivity is a function of the binding degree (number of connections in the database for a given protein) of proteins within the network. Thus, the apparent density of a network could simply be due to the lack of specificity with which its constituents bind in vitro. Second, certain processes are more extensively studied, so more connections between proteins involved in them may be reported (see Figure S1; immune proteins are reported in more publications and have a higher mean binding degree). This confounds our effort to assess connectivity of associated loci if there is a bias in the data. From a genetic standpoint, a common randomization method would involve sampling SNPs from the genome matched for the appropriate parameters (such as gene density and protein binding degree). This method becomes highly limited if the disease loci contain genes that are better studied than the randomly sampled SNPs.
Therefore, we apply a permutation method that is robust to non-specific binding and differences in publication density. We perform a within-degree node-label permutation that is carried out as follows: a random network is built that has nearly the exact same structure as the original InWeb network, only the node labels (i.e. the protein names) are randomly re-assigned to nodes of equal binding degree; this method assumes a null distribution of connectivity that is entirely a function of the binding degree of individual proteins. Random networks will have the same size, number of edges and per-protein binding degree as InWeb; we build 50,000 different random networks. With this method, we are able to test the non-randomness of our network conditional on the exact binding degree distribution of our disease proteins.
Others have used PPI data in complex disease to understand epistatic loci or to build a network of interacting proteins from associated loci [33][34][35]. The novelty of our method lies not in the idea that PPI data can be used to help understand genetic loci associated to disease, but rather in that we have developed a broadly-applicable method to statistically evaluate the degree to which non-random PPI networks emerge from loci associated to complex disease and to leverage from this insight about causal proteins in large loci [33,34]. We show this to be the case in a number of diseases.
Here, we use this methodology to evaluate whether genes in loci associated to five complex traits are significantly connected via protein-protein interactions. We report an algorithm to build and assess PPI networks using the InWeb database and find robust, statistically significant networks underlying associations to RA, CD, height and lipid levels, which we suggest as representative of the underlying pathogenic molecular processes. We then perform several detailed analyses on the RA and CD networks to confirm that they contain true biological insight into disease. We use independent mRNA expression data to show that the prioritized associated proteins we propose as interacting are co-expressed in relevant immune tissues, supporting a plausible biological setting for our findings as well as the validity of the reported protein-

Author Summary
Genome-wide association studies have uncovered hundreds of DNA changes associated with complex disease. The ultimate promise of these studies is the understanding of disease biology; this goal, however, is not easily achieved because each disease has yielded numerous associations, each one pointing to a region of the genome, rather than a specific causal mutation. Presumably, the causal variants affect components of common molecular processes, and a first step in understanding the disease biology perturbed in patients is to identify connections among regions associated to disease. Since it has been reported in numerous Mendelian diseases that protein products of causal genes tend to physically bind each other, we chose to approach this problem using known protein-protein interactions to test whether any of the products of genes in five complex trait-associated loci bind each other. We applied several permutation methods and find robustly significant connectivity within four of the traits. In Crohn's disease and rheumatoid arthritis, we are able to show that these genes are co-expressed and that other proteins emerging in the network are enriched for association to disease. These findings suggest that, for the complex traits studied here, associated loci contain variants that affect common molecular processes, rather than distinct mechanisms specific to each association.
protein interactions. Lastly, by analyzing more recent GWAS meta-analysis results, we show that these networks contain components that show significant evidence of further genetic associations: proteins interacting with multiple associated network members and encoded elsewhere in the genome themselves carry an excess of association to disease in the latest meta-analyses of each of these diseases. Our method, available for download, generates an experimentally tractable hypothesis of the molecular underpinnings of pathogenesis.

Network Construction and Evaluation Pipeline
We construct and evaluate networks of disease loci as outlined in Figure 1. We first define associated proteins as gene products encoded in genomic loci harboring variants associated to disease ( Figure 1A, 1B; see Materials and Methods for locus definition). We construct networks of protein-protein interactions representing proteins as nodes connected by an edge if there is in vitro evidence of interaction (InWeb high-confidence interaction set). We build direct networks, in which any two associated proteins can be connected by exactly one edge, and indirect networks, where associated proteins can be connected via common interactor proteins (not encoded in associated loci) with which the associated proteins each share an edge. We restrict direct and indirect interactions to only those between proteins encoded in distinct associated loci.
We then calculate several metrics to evaluate network properties. These metrics can be divided into two categories: an edge metric and node metrics. The edge metric is the direct network connectivity parameter defined as the number of edges in the direct network. We interpret direct network connectivity as the frequency with which different loci harbor proteins that directly bind each other, regardless of how they assemble; direct network connectivity is therefore our most straightforward metric. Node metrics include the following: associated protein direct connectivity and associated protein indirect connectivity which refer to the number of distinct loci an A. Genes overlapping the wingspan of associated SNPs are defined, and these genes code for associated proteins. B. Associated proteins are used to recover direct and indirect networks. Direct networks (left) are built from direct interactions between associated proteins according to the InWeb database (colored proteins). Connections between proteins within the same locus are not considered. Indirect networks (right) are built by allowing connections between associated proteins through a protein elsewhere in the genome (grey). Various network parameters to quantify connectivity, defined in the text, are assigned. C. Random networks are built from a within-degree node-label permutation method described in Text S1. An empirical distribution is constructed for each network parameter and used to evaluate the significance of networks. D. Using the same permutation method to score individual proteins, a subset of proteins per locus is nominated as candidates for harboring causal variants (red circles). Scores used to nominate candidates, described in Text S1, are Bonferroni corrected for the number of possible candidates within each locus. E. Candidate genes from D (nominal p-values used) are tested for co-expression. doi:10.1371/journal.pgen.1001273.g001 associated protein can be connected to directly and indirectly, respectively, and common interactor connectivity which refers to the average number of proteins in distinct loci bound by common interactors in indirect networks. We interpret all three node metrics as descriptive of the type of network that was constructed: a stream of connections (such as the network A-B-C-D-E) will likely have low and insignificant node metrics despite a significant edge metric, whereas a more tightly clustered network might be enriched for both edge and node metrics. We assess the statistical significance of the various connectivity parameters using a withindegree node-label permutation strategy that controls for variation in the degree to which certain proteins are studied or behave in vitro ( Figure 1C; see Text S1 for details on the permutation strategy, evaluation of its ability to distinguish signal from noise and a benchmark analysis of Fanconi Anemia). As we are interested in the processes underlying disease, we also define the gene encoding the top-scoring protein in each locus as most likely to be causal for association ( Figure 1D; see Text S1 for prioritization strategy). We then use tissue expression data to test whether our nominated candidate genes are enriched in the same tissue(s) and therefore participate in a network that is biologically feasible ( Figure 1E; Text S1). With this approach, we aim to construct plausible models of biological networks underlying pathogenesis.
Our approach controls for biases in the data: using the highconfidence interactions from InWeb addresses laboratory artifacts, and node-label permutation accounts for ascertainment biases due to differing levels of knowledge on biological processes for those proteins present in InWeb ( Figure S1). We show empirically that priority scores given to proteins have no correlation with the degree to which they are represented in the database ( Figure S2). A fundamental limitation of any functional data is that genes for which data are missing will not be considered. This applies to similar methods, including expression data that can be limited to genes represented on specific arrays or ontology analyses that are restricted to well characterized genes. Here, proteins that are entirely absent from the filtered InWeb data are not considered in our analysis (see Discussion). It is important to note that these genes, listed in Table S1, cannot be ruled out as potentially affected by causal variation since we have no power to make such a conclusion. We note, however, that the loci we have considered here (for the 5 complex traits) have the majority of their genes present in the high-confidence InWeb database (Table S1, median inclusion of 81.5%).
We also tested two additional permutation strategies on RA and CD -one based on random sampling of SNPs from the genome matched for proximal gene content and protein binding degree and the other based on edge permutation -that generally provided equivalent results ( Figures S3 and S4); however SNP permutation may not be robust in the presence of extremes of gene density or protein binding degree at some loci, and edge permutation does not preserve the network structure of InWeb (Text S1). This analysis pipeline, which we call Disease Association Protein-Protein Link Evaluator (DAPPLE), is available for download at http://www.broadinstitute.org/,rossin/DAPPLE.

Gene Products Encoded in Associated Loci Interact
We first tested the method on the Mendelian disease Fanconi Anemia (FA) as a proof of principle. We input 9 of the FA genes and found 23 connections among them; compared to 50,000 random networks, the FA network is enriched for connectivity (direct network connectivity p,,2610 25 , Figure S5, Text S1). This result is consistent with current understanding of how the FA genes code for proteins that are part of the same DNA repair complex [36].
We then set out to test our method on two autoimmune diseases that are both complex traits. Recent GWA studies in autoimmune and inflammatory diseases have been particularly successful at determining loci encoding risk to disease, with over 100 loci described to date [2][3][4][5][6][7]1]. We investigated rheumatoid arthritis (RA) and Crohn's disease (CD) and built networks from proteins encoded in 25 and 27 gene-containing associated loci, respectively [2,8]. As described above, we built direct and indirect networks for each set of loci, evaluated the significance of the 4 network metrics to assess the probability that such networks could arise by chance, and we nominated candidate genes by assessing network participation. We followed up our results by assessing tissue coexpression as a test for the biological feasibility.
We were able to connect 20/27 loci for RA and 12/25 loci for CD in direct networks, strongly suggesting functional connections between proteins encoded in the associated regions. When compared to 50,000 random networks, we found that the direct network connectivity (the number of direct network edges) was statistically significant (27 for each disease; P RA = 3610 24 , P CD = 1.11610 23 ; Figure 2) as was the associated protein direct connectivity ( Figure S6A and S6B, P RA = 0.02, P CD = 0.00305). Thus disease-associated loci encode directly interacting proteins beyond chance expectation, suggesting that risk variants may act on suites of proteins involved in the same process.
We were then able to connect all but one gene-containing associated loci in each disease by expanding our networks to include common interactors (26/27 in RA; 24/25 in CD). The associated protein indirect connectivity was significantly enriched in both diseases ( Figure S6A and S6B p = 1610 25 in RA, p = 4.1610 24 for CD), as was the common interactor connectivity ( Figure S6A and S6B, p = 7610 25 for RA and p = 1.1610 24 for CD).
In aggregate, these results suggest that the observations of connectivity in Mendelian diseases are recapitulated in both RA and CD and that common risk variants predisposing to these diseases may impact sets of interacting proteins.
Given the significant connectivity of common interactors in the indirect networks for RA and CD, we speculated that common interactors might themselves be affected by previously undescribed risk variation. To test this, we consulted association data for each disease in the available data from meta-analyses, which for RA was in a newly completed meta-analysis and for CD was the same study that yielded the 30 loci [2,37]. We assigned each recombination hotspot-bounded linkage-disequilibrium (LD) block in the genome an association score that represents the maximum score in that block corrected for the number independent SNPs therein. Genes were assigned association scores based on the blocks they overlap; this score distribution can then be compared to the scores of all gene-containing blocks in the genome (for both diseases, we removed the MHC from this analysis due to LD properties). Using this method, we found that common interactors expressed in the same tissues as associated proteins in our networks (see below) were encoded in regions with evidence of association significantly in excess to what is expected in gene-containing regions. In RA, the distribution of common interactor scores was skewed toward higher association (one-tailed rank sum p = 1.7610 25 ) and in CD, we saw similar enrichment (p = 6.5610 24 ). See Text S1 for details of analysis. This observed skew suggests that the common interactors themselves may harbor risk variants; we therefore considered the regions they overlap as candidates for replication (see ''Crohn's Network Predicts New Loci'' section).

Extending Analysis to Height, Lipids, and Type 2 Diabetes
To test whether the observed significant connectivity seen in RA and CD was present in non-immune complex traits, we tested our method on three traits: human height, blood lipid concentration (both LDL and HDL) and Type 2 Diabetes (T2D). We used 37 replicated gene-containing loci associated with human height, 18 with blood lipid levels and 36 with T2D [9][10][11][12][13][14][15][16][17]. The loci associated to height and lipids each contain proteins that assemble into significantly connected direct networks ( Figure S6C and S6D, direct network connectivity p = 1610 24 and p = 1.9610 24 for each disease, respectively; see Text S1 for significance of other 3 parameters). In the height network, 19/42 loci participated in the direct network and 34/42 participate in the indirect networks, but only the direct network connectivity and the common interactor connectivity were significantly greater than chance. In the lipids network, 11/ 19 participated in the direct network and 16/19 in the indirect; all node metrics except the common interactor connectivity were significantly enriched. 9/37 T2D loci participated in the direct network and 34/37 in the indirect network; however, 3/4 metrics were not greater than chance expectation and only one was slightly enriched ( Figure S6E, network connectivity p = 0.44960; see Text S1 for significance of other 3 parameters).
We therefore conclude that the PPI connectivity seen in two autoimmune diseases can be generalized to other complex trait loci (height-and lipid-associated regions), though we could not confirm the significance of the T2D network.
Our results suggest that functionally connected proteins reside in regions of the genome associated to disease risk. Permutation analysis revealed that these connections are in excess compared to what is expected given the binding profiles of associated proteins. For RA and CD, other proteins interacting with the associated proteins also show evidence of association beyond chance expectation. Cumulatively, these findings suggest that risk to the complex disease/traits studied here is spread over functional groups of proteins, directly analogous to observations in Mendelian traits.

Prioritizing Proteins in Associated Loci Reveals Likely Pathogenic Tissues
An obstacle to interpreting GWA results stems from the difficulty in identifying the genes within associated regions influenced by risk variants. Candidate genes are often selected based on proximity to most associated markers and miscellaneous forms of previous knowledge. We therefore asked whether our observations could lead us to a principled, data-driven approach to selecting candidate genes by assessing their role in our networks. As shown in Figure 1 and described in detail in the Text S1, we used an iterative optimization method to assign priority scores to associated genes based on the network participation of their encoded proteins. We nominate genes that achieve the best score within their locus as the candidates for influencing disease risk. We describe the results in detail here for RA and CD; see Table S2 for scores assigned to RA, CD, blood lipid level and height genes.
We were able to nominate candidate genes in 12/21 RA loci encoding multiple genes (Table S2; Text S1). Examples of candidate genes in RA were IL2RA, CD40, CD28, PTPN22, CTLA4 and TRAF1. We accomplished the same task in CD, nominating candidate genes in 10/18 multi-genic loci. Candidates included JAK2, STAT3, IL23R/IL12RB2, PTPN2, MST1R and AIRE. For both diseases, genes in single-gene loci are also scored, though they are automatically considered the candidate gene (but not necessarily part of the underlying mechanism). It is important to note that we do not expect high-scoring proteins in every locus; we only expect high scores for those proteins that may participate in the common process(es) detected via enrichment in connections. RA and CD, like most complex diseases, most likely have many underlying processes, not all of which are captured here.
The core networks involving only these candidate genes represent our mechanistic predictions of pathways underlying pathogenesis in RA and CD. From a statistical standpoint the final networks built from candidate proteins account for the excess connectivity that we initially observed: the significance remains if we restrict multi-genic loci to just these genes ( Figure S7A-S7D, direct network connectivity p,2610 25 for RA and CD), while networks built from the remaining non-prioritized genes are less significant ( Figure S8, direct network connectivity p = 0.0368 and p = 0.993, for RA and CD respectively). The remaining significance in RA is most likely a sign of additional important proteins that did not make the cutoff. From a biological standpoint, our candidates agree with experimental findings in the few cases where such evidence exists [38][39][40][41][42][43][44][45]. We therefore show that the connectivity between associated loci in RA and CD is primarily driven by a small subset of associated proteins encoded in those regions; this observation suggests that the interacting proteinsand the biological pathways they represent -may be the targets of risk variation.
To test the biological plausibility of our nominated core networks, we asked whether the candidate genes are co-enriched in subsets of particularly relevant tissues in a reference microarray dataset consisting of 14,184 transcripts measured in 55 immune, 8 gastro-intestinal, 27 neurological and 36 miscellaneous other tissues (126 total) [46]. These publicly available data are curated: expression intensities were converted to enrichment scores to reflect the enrichment of a gene in a tissue given its expression in all tissues. For each tissue, we compared the expression enrichment of RA and CD candidate genes to the rest of the genes in the genome using a one-tailed rank-sum test, resulting in a p-value for each tissue. A significant difference for a given tissue indicated that the genes in question were enriched for expression in that tissue compared to all genes in the genome. We also performed the same analysis for the remaining non-prioritized genes in associated regions to test whether the network prioritization method identified genes that were enriched in tissues distinct from non-prioritized genes. For discussion purposes, we defined ''top'' tissues as tissues achieving p,0.1 ( Figure 3 depicts the entire distribution of p-values). This analysis led to 3 main conclusions. First, we found that for each disease, enrichment only occurred in immunologically relevant tissues (Figure 3; strikingly, immune tissues are nearly all ranked higher than other tissues). Second, we found that this was not the case to such an extent for non-prioritized genes (Figure 3, black points). Third, we found that the non-prioritized genes had fewer tissues where we could detect enrichment (Figure 3, RA and CD candidate gene tissue scores are more significant than tissue scores of non-prioritized genes). We formally tested this by comparing the p-value distributions for candidate genes and non-prioritized genes using a one-tailed ranksum test (p = 2.85610 27 for RA; p = 2.55610 24 for CD). Of the 11 top tissues for CD candidate genes, 7 are subgroups of T-cell lymphocytes; the analogous list for RA (21 tissues) contains a mix of immune tissues, again dominated by T-cell subgroups ( Table 1). The top tissue compartment for both diseases is defined as CD4+ T-cells.

Crohn's Network Predicts New Loci
We hypothesized that a subset of proteins connected to the core CD network ( Figure 4B, the network built from prioritized genes in CD loci) might be near true causal variation. Having observed significant enrichment for association in the common interactors, we then chose a more conservative approach to propose candidate genes. We selected all proteins that connect directly to the core CD network only (21 genes) and filtered them on expression in the relevant tissues (Table 1). While this manuscript was being prepared, a larger meta-analysis was completed and recently published that reports 39 new loci associated to CD (295 overlapping genes) [47]. Of the 293 genes proposed by our method (small circles, Figure 4B), 10 were in newly associated regions (small red circles). This represents a statistically significant enrichment compared to chance expectation based on random draws from all 21,718 genes (p = 0.001) as well as random draws from genes expressed in at least one of the CD-relevant tissues (p = 0.01). . Candidate RA and CD genes are preferentially expressed in immune tissues. We obtained tissue expression data for 126 different cell types from a publicly available database, which was grouped into immune, gastrointestinal (GI), neuronal and 'other' [46]. For each tissue, we compared the expression of RA (A) and CD (B) candidate genes to the rest of the genes in the genome using a onetailed rank-sum test, resulting in a p-value for each tissue (-log(p) is plotted on the y-axis). A significant difference for a given tissue indicated that the candidate genes were enriched for expression in that tissue compared to all genes in the genome. To test whether our network prioritization identified genes that were co-enriched in specific tissues beyond what was expected from all genes in associated regions, we calculated the same p-values for the rest of the genes in RA and CD associated loci (i.e., the genes that weren't prioritized via our network permutations). In this figure, we plot the tissue enrichment scores for each tissue for the candidate genes (purple) and the non-prioritized genes in the remaining regions of association (black). We indicate the category of tissue on the bottom: immune (red), GI (yellow), neuronal (green) and other (blue). We ordered the tissues by decreasing enrichment score of the candidate genes. doi:10.1371/journal.pgen.1001273.g003 We performed a similar analysis in RA since the recent metaanalysis discovered 6 new loci (18 new genes) [37]. Of the 610 genes proposed, 1 was among the 18 new genes ( Figure 4A, small red circle). This does not represent a statistically significant enrichment.

Candidate Gene Networks Suggest Underlying Biology
The networks (Figure 4) suggest pathogenic mechanisms in agreement with current thinking on disease etiology and propose novel roles for candidate proteins in these pathways. The RA network ( Figure 4A) appears to represent signaling cascades involved in the inhibition or stimulation of the NF-kB complex, a factor that activates transcription of genes encoding cytokines, antibodies, costimulatory molecules and surface receptors [43]. STAT4 encodes a transcription factor that is activated upon engagement of cytokines, such as IL12 and interferon type I, with their receptors [43]. We show that not only does STAT4 show enrichment for connectivity, it is connected indirectly to a number of associated genes encoding surface receptor subunits that also achieve high network scores, such as IL12RB, IL2RA and PTPRC. TNFAIP3 (known as A20 in mice) is a cytoplasmic zinc finger protein that inhibits NF-kB activity, and knockout mice develop widespread and ultimately lethal inflammation, making it a plausible player in RA pathogenesis [48]. Also in the NF-kB pathway is associated protein CD40, which scores highly in our networks and binds TRAF6 and TRAF1 directly. CD40 is normally found on B cells but has also been shown to act as a costimulatory molecule on T cells to augment CD28 response and activate NF-kB [49].
PTPN22, a gene with strong genetic support for harboring risk variants (including the strongly associated R620W coding polymorphism), has been shown to act as a negative regulator of TCR but has not yet been definitively linked to a pathogenic mechanism [43,50]. Here, we place it in context of other highly associated proteins and suggest that it is part of a common mechanism.
Finally, the RA network places a number of other proteins that have not yet been formally studied in the context of the proposed network underlying RA; these include CD2 and CD48, as well as FCG2RA and PRKCQ, genes suspected of being causal but not formally placed in a mechanism with other associations.
In CD the core of the candidate network (IL12B/IL23R/ JAK2/STAT3; Figure 4B) corresponds to the interleukin-23 (IL23) signaling pathway. IL12B encodes p40, a component of the heterodimeric IL23. The IL23R gene encodes one half of the also heterodimeric IL23 receptor. This receptor is a cell surface complex found on a variety of immune cells; on activation, it induces Janus Kinase 2 (Jak2) autophosphorylation, which in turn leads to the translocation of STAT3 to the nucleus to activate transcription of various pro-inflammatory cytokines [40]. IL23 signaling is necessary for the activation and maintenance of a subset of CD4 + T cells acting as 'inflammatory effectors'; these interleukin-17 responsive T-cells (Th 17 ) have been implicated in autoimmune inflammation in CD and experimental models of other autoimmune diseases [40]. We note that IL23 belongs to the interleukin 12 family of cytokines and both ligand and receptor share subunits with the canonical IL12-mediated signaling pathway, which induces activation of regulatory T cells (T reg ). Expression data was downloaded from a publically available dataset [46]. The data had been previously converted into enrichment scores (see Materials and Methods). The enrichment scores of candidate genes in RA and CD were compared to the rest of the genome by a one-tailed rank-sum test. The tissues that received a p-value of ,0.1 are shown. Of note, all tissues in this category for both RA and CD are immune, as shown in Figure 3. doi:10.1371/journal.pgen.1001273.t001 Our CD network suggests that other proteins participate in this pathway, including the tyrosine phosphatases encoded by PTPN2, a gene also associated to other autoimmune diseases [51]. Other proteins that are indirectly connected to this pathway include IRF1, which we score highly and that has separately been reported to activate transcription of IL12RB1 [52]. Furthermore, the common interactors that we prioritize for replication of association given their involvement in the CD network -including JAK1, STAT4, TYK2 and IL2RA -fall into the IL12 and IL23 signaling pathway (TYK2 and IL2RA were of the genes recently found to be in regions of association).
The CD network also generates new hypotheses about potentially important genes. We prioritize AIRE, an associated protein involved in T-cell development, which has not been extensively studied in the context of Crohn's but could plausibly lead to autoimmunity. ZNF365, a gene that achieves a high permutation score, has been assumed to be the causal gene because it is the only gene to reside in the wingspan of its locus; however, it has not been studied as part of the core network described here (IL23R/JAK2/STAT3 pathway). Finally, CSF2, IKZF3 and GRB7 are in the same large locus (17 genes) but achieve significant permutation scores; these genes have been less well studied in the context of CD.

Discussion
We have shown that proteins encoded in regions associated to RA, CD, height and lipids interact and that the networks they form are significantly connected when compared to random networks. In CD and RA, the genes encoding prioritized proteins are preferentially expressed in immune tissues relevant to the pathogenesis of both diseases, while the rest of the genes in associated loci show less tissue preference. Furthermore, we can connect other associated proteins to these networks via common interactors, which appear to be encoded in genomic regions harboring further risk variants. Newly available data in CD allowed us to confirm that genes predicted to be near causal variation are indeed in regions now known to be associated to CD. We note that the conclusion of connectivity could not be extended to T2D, and we hypothesize that the lack of connectivity may be due to disparate underlying mechanisms that have yet to be well captured genetically. Though our aim was to build and analyze networks that emerge from replicated regions of association, we feel that a promising future direction may be to look more broadly for networks enriched in weaker signals of association. Evidence that this type of analysis may be helpful is that we pointed to a set of weaker CD association signals that were found to be true positives in a larger study.
Our results have several implications for the interpretation of genome-wide association studies: first, our ability to connect the majority of associated loci in a limited number of molecular networks suggests that these represent processes underlying pathogenesis. Second, these networks are unbiased, in the sense that they do not rely on previous classifications of gene function or pathway lists; rather, we assemble our networks from low-level functional genomics data and allow network structure, if any, to emerge. Third, our approach is general; we have demonstrated it using interactions between protein products, but any relationship between genes or other genomic features (non-coding RNAs, enhancer elements, conserved regions etc.) may be used in the same fashion. Even more powerful, approaches combining such orthogonal data types will be rewarding. The limitation to using PPI data from a curated database such as InWeb is that proteins for which no high-confidence interactions exist will be left out of the analysis. As such, our analysis is limited to proteins present in the database. Additionally, while we controlled for the biases we observed, other undetected biases still may exist. . Final disease networks. Resultant networks built from candidate genes are depicted for RA and CD (A and B, respectively). Using only the candidate genes, we plotted the direct network as well as any other proteins connected to the direct network after filtering them on expression in any one of the tissues found to be specific to the core network. 610 such proteins connect to the RA network and 293 such proteins connect to the CD network. Large circles represent disease proteins, and small circles represent the connected proteins. Small red circles indicate proteins connected to the core network that were newly identified associated regions (10 proteins in CD and 1 protein in RA). The large circles are colored by locus. doi:10.1371/journal.pgen.1001273.g004 Interestingly, there are certain cases where the method is able to distinguish between proteins that are close in the genome and functionally very similar. In RA, the rs12746613 locus has 3 genes in the PPI database -FCG2RA, FCGR3A and HSP70B. FCG2RA achieved a nominal p-value of 0.00703, whereas FCGR3A achieved p = 0.38296. Similarly, in the large rs3197999 locus in CD, the method gave MST1R a p-value of 0.0066 whereas MST1, the ligand of MST1R, achieved a p-value of 1. In these cases, the method is able to distinguish between functionally similar genes. There are times when it is unable to distinguish between functionally similar genes, however, such as the IL21/IL2 locus in RA, the STAT1/STAT4 locus in RA and the STAT3/STAT5A/ STAT5B locus in CD.
We note in passing that the candidate genes we nominate are on average the closest to the most associated SNP in each locus, even though proximity within the LD region was not considered in the PPI analysis (p = 0.005, Figure S9). This supports the theory that the majority of causal variation will be close to the association signal rather than anywhere in the region of LD. We also observed overlap between genes prioritized by this method and GRAIL, a text-mining approach that uses orthogonal data (Table S2) [19]. We depict this information, as well as overlap between prioritized genes and the presence of non-synonymous SNPs, in Figure S10.
In this paper, we have studied 5 complex phenotypes, 4 of which show evidence of abundant PPI connections across loci. Our results therefore allow us to speculate that other complex diseases may behave in the same way and that genetic risk may be spread over the molecular processes that influence disease, rather than a single, catastrophic mutation as in Mendelian inheritance. In order to determine whether what we find here is expandable to complex disease in general, however, we would need to apply our method to the many more diseases and traits to which regions of the genome have been associated. Nonetheless, for the networks that emerge here, our approach identifies sets of proteins plausibly involved in pathogenesis, and the next step will be to identify what the molecular and phenotypic consequences of perturbing such processes are and how they relate to overall disease etiology.

InWeb Database
We used a probabilistic database of reported protein-protein interactions described in 2007 by Lage et al [24,31]. This database contains 428,430 reported interactions, 169,810 of which are deemed high-confidence, non-self interactions across 12,793 proteins. High-confidence is defined by a rigorously tested signal to noise threshold as determined by comparison to well-established interactions [24]. Briefly, InWeb combines reported protein interactions from MINT, BIND, IntAct, KEGG annotated protein-protein interactions (PPrel), KEGG Enzymes involved in neighboring steps (ECrel), Reactome and others as described elsewhere in detail [53][54][55][56][57][58][59][60][61]. All human interactions were pooled and interactions in orthologous protein pairs passing a strict threshold for orthology were included. Each interaction was assigned a probabilistic score based on the neighborhood of the interaction, the scale of the experiment in which the interaction was reported and the number of different publications in which the interaction had been sited. The data we used is available at www. broadinstitute.org/,rossin/PPI/ppi.html.

Disease Loci
30 CD SNPs were derived from the first CD meta analysis of which 25 contain genes [2]. 28 RA SNPs were derived from the most recent RA review of which 27 contain genes [8]. 42 Height SNPs were derived from a number of analyses of which 38 contain genes [10,13,16]. 19 blood lipid level SNPs were derived from a number of analyses of which all 19 contain genes [11,15]. Finally, 42 T2D SNPs were derived from a number of analyses of which 37 contain genes [9,12,14,17].

Translating SNPs to Genes
Hotspot and linkage disequilibrium (LD) information were downloaded from www.hapmap.org for CEU hg17 and hg18 to match the version in which associations were reported [62]. We defined the wingspan of a SNP as the region containing SNPs with r 2 .0.5 to the associated SNP; this region is then extended to the nearest recombination hotspot. We downloaded the Ensembl human gene list from UCSC Genome Browser and collapsed isoforms into single genes [63]. We converted gene IDs from Ensemble to InWeb IDs. A gene's residence in a locus is defined by whether 110 kb upstream and 40 kb downstream (to include regulatory DNA) of the coding region of the gene's largest isoform overlaps the SNP wingspan [64].

Statistical Analysis
All analyses, including building networks and evaluating significance, were carried out in R, Perl and Python and are available at www.broadinstitute.org/,rossin/PPI/ppi.html. Details on the algorithms are available in the Text S1 file. Figure S1 Immune proteins are more likely to be in publications involving reported protein-protein interactions. For each protein, we enumerated the number of unique publications in which it is reported. We then categorized proteins as ''immune'' or ''nonimmune'' based on the same expression data used in Figure 3. Genes in the top 2% of expression for at least 2 of the 55 immune tissues were deemed immune genes; their proteins were then deemed immune-proteins. The distribution of publications for immune proteins is significantly different than the background distribution (Kolmogorov-Smirnov p-value ,2E-16). Found at: doi:10.1371/journal.pgen.1001273.s001 (0.13 MB TIF) Figure S2 Correlation between prioritization p-value and binding degree. We show that there is no correlation between prioritization p-values given to genes and their representation in the database. We show this in randomized networks (A) and in the disease networks (B). Found at: doi:10.1371/journal.pgen.1001273.s002 (0.31 MB TIF) Figure S3 RA and CD networks are significantly connected when tested by SNP matching. As a test for robust enrichment in connectivity, we built 1,000 networks from SNPs matched for gene density throughout the genome for both RA and CD. These networks were then filtered for SNPs whose associated genes have a comparable binding degree distribution to those in RA and CD (847 remain in RA, 353 remain in CD, see Text S1). We show the expected random distribution in relation to the disease network for each of the 4 parameters (direct network connectivity, associated protein direct connectivity, associated protein indirect connectivity, common interactor connectivity) for RA (A) and CD (B). Permuted networks were generated as described in Text S1. Results found by node-label permutation were similar, though less so for RA.  Figure S4 RA and CD networks are significantly connected when tested by edge shuffling. As a test for robust enrichment in connectivity, we built 1,000 networks using an edge shuffling method (see Text S1 for details). We show the expected random distribution in relation to the disease network for each of the 4 parameters (direct network connectivity, associated protein direct connectivity, associated protein indirect connectivity, common interactor connectivity) for RA (A) and CD (B). Permuted networks were generated as described in Text S1. Results found by node-label permutation were similar. From left to right and top to bottom for RA, p,0.001, = 0.013,,.001, and ,0.001. From left to right and top to bottom for CD, p = ,0.001, ,0.001, ,0.001, and = 0.001. Found at: doi:10.1371/journal.pgen.1001273.s004 (0.70 MB PDF) Figure S5 Fanconia Anemia network. As a benchmark analysis, we tested the method on Fanconia Anemia, a Mendelian disorder known to be caused by genes coding for interacting proteins. 9 of the 13 FA genes were in the InWeb database. We found that the direct network connectivity was 23, which is many more than expected by chance (p,,2E-5). The associated protein direct connectivity, associated protein indirect connectivity and common interactor connectivity were all significantly enriched (p,2E-5, p = 0.004, p = 0.009, respectively). These results agree with the current understanding of FA pathogenesis. FA Network is shown.  Figure S7 RA and CD networks built from candidate genes are significantly connected. Using the permutation method, we nominated candidate genes as those that are more highly connected to the network than chance expectation. If such a high-scoring protein is found, the remaining proteins in the locus scoring p.0.05 are removed. The remaining network is significantly enriched for connectivity. We show the null distribution in relation to the disease network for each of the 4 parameters (direct network connectivity, associated protein direct connectivity, associated protein indirect connectivity, common interactor connectivity) for RA (A) and CD (B). From left to right and top to bottom for RA, p,2E-5, = 0.017, ,2E-5 and ,2E-5. From left to right and top to bottom for CD, p,2E-5, = 0.008, ,2E-5 and ,2E-5. Found at: doi:10.1371/journal.pgen.1001273.s007 (0.70 MB PDF) Figure S8 The remaining non-candidate proteins are not enriched for connectivity in RA and CD. Those proteins that are filtered out during the candidate gene process are not enriched for connectivity. We show the expected random distribution in relation to the disease network for each of the 4 parameters (direct network connectivity, associated protein direct connectivity, associated protein indirect connectivity, common interactor connectivity) for RA (A) and CD (B). From left to right and top to bottom for RA, p = 0.0368, 0.2424, 0.7274 and 0.1748. From left to right and top to bottom for CD, p = 0.993, 0.999, 0.6558 and 0.7468. Found at: doi:10.1371/journal.pgen.1001273.s008 (0.72 MB PDF) Figure S9 Candidate genes are likely to be near to the associated SNP. Candidate genes within multigenic loci were prioritized as described. We defined the distance from a gene to the SNP that tags it as the shortest of two distances: the distance from its start codon to the SNP and its stop codon to the SNP. Genes within a SNP's wingspan are then given ranks as to how close they are to the SNP (closest gene, 2nd closest gene, and so on). These distances were collected for RA and CD and the distribution is shown (salmon bars). We compared this distribution to the distribution of 100 simulated distances as defined by random assignment of candidate genes in associated loci (black hatched bars). The distributions are significantly different (one-tailed Kolmogorov-Smirnov test p = 0.008). Found at: doi:10.1371/journal.pgen.1001273.s009 (0.08 MB TIF) Figure S10 Overlap of prioritized genes across methods. For each SNP, we compared the prioritized genes through PPI networks, through GRAIL and through presence of nonsynonymous SNPs. We show the overlap for all SNPs, where two methods agree if at least one prioritized gene in the region is the same. Found at: doi:10.1371/journal.pgen.1001273.s010 (0.02 MB PDF)

Table S1
Enumerating genes included and missing from the analysis. For each disease, each locus (column 2) is annotated with the number of genes in the locus according to the Ensembl gene list downloaded from the UCSC Genome Browser (column 3 and 6) and the number of genes in the locus for which InWeb highconfidence binding data is available (column 4 and 7). The percent of genes included in the analysis out of the total possible genes was calculated (column 5). The median gene inclusion is 81.5%. *Genes in the Ensembl gene list that are not protein-coding or proven to be protein-coding (rna genes, pseudogenes, theoretical proteins) are not included in the ''All Genes'' list. Found at: doi:10.1371/journal.pgen.1001273.s011 (0.05 MB XLS) Table S2 CD and RA candidate genes proposed through permutation. This file provides the candidate gene scores for all CD and RA genes involved in the networks (A), a comparison of prioritized genes to those of GRAIL and genes with coding SNPs for Crohn's disease (B), and a full comparison of GRAIL p-values for all 4 significantly connected traits (C).