• We are sorry, but NCBI web applications do not support your browser and may not function properly. More information
Logo of nihpaAbout Author manuscriptsSubmit a manuscriptNIH Public Access; Author Manuscript; Accepted for publication in peer reviewed journal;
Genomics. Author manuscript; available in PMC Nov 1, 2009.
Published in final edited form as:
PMCID: PMC2602835
NIHMSID: NIHMS76570

Pathway Analysis of Seven Common Diseases Assessed by Genome-Wide Association

Abstract

Recent genome wide association studies (GWAS) have identified DNA sequence variations that exhibit unequivocal statistical associations with many common chronic diseases. However, the vast majority of these studies identified variations that explain only a very small fraction of disease burden in the population at large, suggesting that other factors, such as multiple rare or low-penetrance variations and interacting environmental factors, are major contributors to disease susceptibility. Identifying multiple low penetrance variations (or ‘polygenes’) contributing to disease susceptibility will be difficult. We present a pathway analysis approach to characterizing the likely polygenic basis of seven common diseases using the Wellcome Trust Case Control Consortium (WTCCC) GWAS results. We identify numerous pathways implicated in disease predisposition that would have not been revealed using standard single-locus GWAS statistical analysis criteria. Many of these pathways have long been assumed to contain polymorphic genes that lead to disease predisposition. Additionally, we analyze the genetic relationships between the seven diseases, and based upon similarities with respect to the associated genes and pathways affected in each, propose a new way of categorizing the diseases.

Keywords: Pathway, genome-wide, disease, common, diabetes, crohn’s, coronary, bipolar, arthritis, hypertension

Introduction

Initial results from single locus-based GWAS analyses suggest that only a minority of DNA sequence variations that modulate disease susceptibility will be identifiable within the GWAS framework. The reasons for this are numerous and likely reflect the fact that most common chronic diseases are influenced by a number of rare and/or low penetrance variations that interact with environmental factors in complicated ways [1]. Thus, the very small fraction of both the heritable component and population disease burden explained by the few polymorphisms identified as statistically significant in recent GWAS initiatives indicates that a large number of risk alleles – the vast majority of which are not likely to have odds ratios greater than 1.3 – are effectively invisible to current single-locus GWAS analysis methods [1,2].

However, it is likely that alternative analysis approaches to GWAS data that focus on the combined effects of many loci, each making a small contribution to overall disease susceptibility, may reveal insights into the genetic basis of common chronic disease. For example, it is highly probable that univariate, single locus analysis results contain informative trends that, when viewed in the contexts of genetic networks and fundamental molecular pathways, can expose aspects of the polygenic basis of disease susceptibility. Although a few methods have been proposed for detecting gene-gene interactions in GWAS initiatives [3,4,5], both their use and comprehensiveness is limited. In addition, it is not yet known whether overt interaction effects contribute as much to the expression of a phenotype as the combined, additive effects of many variations each with small individual (i.e., polygenic) effects. The availability of the Wellcome Trust Case Control Consortium (WTCCC) GWAS data and single locus analysis results [6], provides an excellent opportunity to develop and test more comprehensive multi-locus analysis methods designed to tease out aspects of the polygenic basis of disease from GWAS data, and hence may provide insights into the combined effects of low penetrance variations underlying disease susceptibility.

One approach to pathway-based analyses of GWAS data involves converting the data to a form amenable to analyses with tools developed for the interpretation of microarray-based gene expression studies. In this light, we used the pathway and network analysis tools provided in the MetaCore software suite developed by GeneGo (www.genego.com). Essentially, SNPs are assigned to genes based upon their proximity to open reading frames, and gene weights – as determined from single locus association test statistic data arising from a GWAS analysis – are assigned to each gene (see Methods for details). Trends identified from this gene-specific data that correspond, are consistent with, or reveal, particular networks or pathways can then be tested for their statistical significance. The goal of such an analysis is to determine if the variations that are more strongly associated with a disease tend to aggregate or cluster in biological pathways and networks that are of biological relevance to a disease.

Ultimately, we demonstrate that pathway-based approaches are a powerful means of overcoming the limitations imposed by univariate, single locus analysis of GWAS data, and offer a powerful methodology for revealing the polygenic nature of common chronic disease susceptibility. Our analyses considered the WTCCC GWAS data, which includes data from studies of bipolar disorder (BD), coronary artery disease (CAD), Crohn’s disease (CD), hypertension (HT), rheumatoid arthritis (RA), type 1 diabetes (T1D), and type 2 diabetes (T2D). Our analyses of these diseases expose pathways long known to be involved in disease pathogenesis, such as dopamine signaling in hypertension, but for which causative common variations have not yet been identified. Additionally, comparison of pathways and processes implicated in different diseases reveals surprising insights into shared genetic bases underlying seemingly unrelated diseases, and indicates a potential novel class structure in the categorization of these diseases.

Results

Disease Relationships

The correlations between the seven diseases, based upon the negative log of p-values for genes, adjusted for the number of SNPs mapping to each gene, is presented in upper diagonal entries of Table 1 (see the Methods section for details about the calculations). The lower diagonal of Table 1 presents the number of SNPs shared in common among the top ranked 1000, in terms of statistical significance, SNPs per disease. The large number of SNPs shared in common among all the diseases reveals strong relationships between all the diseases considered. For comparison, the number of SNPs expected to be shared in common by maximum likelihood, according to the hypergeometric distribution, is two. Thus, all disease relationships are strongly significant (p<0.0001 based on the hypergeometric distribution). Similar calculations using Pearson correlation coefficients over the SNPs to measure similarity in the findings across the different diseases also suggest the existence of relationships between the diseases. For example, strong correlations between the metabolic disorders CAD and T2D, as well as autoimmune system-related disorders RA and T1D, were observed, as would be expected by their known etiologies. It should be emphasized, however, that these correlations involve SNPs whose test statistics did not reach genome-wide significance levels and hence would have been considered potential false positive results if not viewed in a polygenic context. However, some surprises were also observed. For example, a strong genetic correlation between BD with metabolic disorders CAD and T2D, as well as a strong correlation between HT and CD -- seemingly unrelated diseases -- was observed. These relationships are seen both from an analysis based on correlations between SNP test statistics (upper diagonal) and an analysis based on the fraction of high-ranking SNPs in common for each pair of diseases (bottom diagonal). In fact, the two different measures of the relationships between diseases themselves are correlated (r2 = 0.449127, p = 0.0411 (Pearsons’ correlation); r2 = 0.64, p = 0.0018 (Spearman’s correlation)). This suggests some unexpected links between these diseases, which will be considered further in the analysis of each disease described in isolation below.

Table 1
Relationships Between the Seven Diseases Based on the GWAS Results

Bipolar Disorder

Supplemental Table 1 presents the top 10 disease pathways enriched among the most significant SNPs identified in the bipolar disorder data, which clearly implicate mood and neurological disorder disease pathways. The signaling pathways enriched in bipolar disorder are presented in Table 2. Notably, regulation of dopamine signaling is a significant pathway. Although a great deal of research has explored a direct link between the regulation of dopamine signaling and mood disorders, there has been little success in identifying statistically compelling associations with particular variations in genes involved in dopamine signaling and mood disorders. On the other hand, the successful targeting of these pathways by drugs used to treat bipolar disorder strongly suggests a link between dopamine signaling and bipolar must exist [7]. Proteins identified as significant from the BD analysis within this pathway are the AMPA receptor, GABA-A receptor beta-1 and beta-3 subunits, as well as PP2A. The biological processes enriched in the BD data (Supplemental Table 2) further implicate glutamatergic signaling and nervous system development, and the occurrence of these pathologies within a BD population validate their importance as putative genetic factors implicated in BD pathogenesis.

Table 2
Significantly Overrepresented Pathways for the Bipolar GWAS.

The first significant pathway, heparan sulfate, contains the associated gene EXTL1, a glucosyltransferase, as well as three sulfotransferase proteins, all of which are expressed in the brain: HSST3, HS6ST3, and HS6ST1. Sulfotransferase activity inactivates dopamine by sulfation, is induced in an animal model of psychotic mania, and it is thought that suflotransferase defects may impair clearing of dopamine from the extracellular space surrounding neurons [8]. Additionally, two other sulfotransferases, SULT1A and C18orf4, have been linked to bipolar disorder [9]. Our finding provides more interesting leads for the role of sulfotransferases in mood disorders and their likely effects on dopamine signaling.

The Niacin-HDL metabolism and adrenergic receptor-mediated cytoskeletal remodeling pathways provide an intriguing link between BD and metabolic disorders, as confirmed by the high degree of correlation between the BD and CAD and T2D GWAS results described in Table 1. In addition to the obvious link between HDL metabolism and metabolic syndrome, both alpha and beta adrenergic signaling have been implicated in metabolic syndrome [10]. In fact, a reasonably high prevalence of metabolic syndrome in patients with bipolar disorder has been observed [11], strongly suggesting a genetic relationship between mood disorders and metabolic disorders.

Coronary Artery Disease

Genome wide association studies have had little success in identifying genes underlying coronary artery disease (CAD), although diabetes, obesity, and metabolic syndrome are well known risk factors for CAD. Analysis of enriched processes among the most strongly associated genes in the CAD data (Supplemental Table 3) reveals a great biochemical and physiologic complexity underlying CAD. However, the dominant processes appear to target contractile functions of vascular smooth muscle cells (VSMC), ranging from cytoskeleton rearrangement, cell adhesion and matrix interactions, development of synapses, and neuromuscular junctions, control of neurotransmission and smooth muscle contractions, and calcium signaling. Most of these processes, especially cell adhesion, matrix interaction, and neuronal factors, have been shown to be important for differentiation of VSMC from migratory to contractile cells [12]. This suggests that genetic defects in these processes may inhibit the contractile phenotype of VSMC and favor phenotypic switching to the migratory VSMC phenotype, a key process in atherosclerosis [12].

While the enriched biological processes among the most strongly CAD-associated genes point towards developmental defects underlying CAD, pathway analysis (Table 3) provides evidence for a series of mechanisms in which the Tubby gene plays a central role. Loss of function Tubby mutants lead to obesity in mice [13], strongly implicating this pathway in metabolic aspects of CAD. However, Tubby itself does not appear as a significant gene in our analysis; rather, PLCG2, which modulates insulin signaling through Tubby [14], and G-protein GNG2, which modulates Serotonin signaling through Tubby [15], as well as PRKCH, are among the highest ranking genes in our analysis. The serotonin signaling pathway is also implicated in our pathway analysis of CAD, and the high correlation between BAD with the BD GWAS results, as well as the strong association of depressive symptoms with cardiac events, lends support to the interesting possibility that common genetic mechanisms may underlie the two disease phenotypes [16]. Tubby signaling members PRKCH and PLCG2 are also involved in VEGF signaling, and another associated gene in that pathway, TEK, appears in our analysis, suggesting that a real genetic link between coronary artery disease and defects in angiogenesis may exist [17].

Table 3
Significantly Overrepresented Pathways for the CAD GWAS.

Finally, calcium signaling appears to play a role in CAD. Calcium channel constituents, CACNA1C, CACNA2D3, CACNB4, and CACNG1, are all affected in the CAD data. Calcium and IP3 signaling are interrelated, and within these pathways we observe PLCG2, CAMK2D, as well as MEF2B and MEF2A. MEF2A gene variations have already been implicated in early onset CAD [18], suggesting that these four highly interconnected pathways contain genes that, when perturbed, may mediate CAD.

Crohn’s Disease

CD is an immunologically-mediated inflammatory condition and the results of our pathway analysis of the most strongly associated genes with CD provides insights into specific immunological pathways that might be altered in CD. A number of pathways support the established hypothesis of aberrant T-cell responses in CD, especially defects in MHC-class II signaling through HLA-DRA to T-helper cells (Table 4). In addition, we observe pathway hits in calcium signaling, specifically at RhoA, AKAP6, and calcium channels CACNA1C and CACNA2D3, all of which modulate calcium levels, which, in turn, through calcinuerin signaling, regulate the transcriptional activity of another associated gene in the calcium signaling pathway, NF-AT. NF-AT, is a central element in T-cell differentiation and proliferation [19]. NF-AT binds to DNA co-operatively with other transcription factor hits in both the IL2 and IL3 signaling pathway, namely AP-1 (FOSL2) and OCT1 (POU2F1) [20]. Interferon signaling is also affected by genes involved in interferon alpha/beta regulatory factors: IRF1 and IRF2 and interferon induced protein IFI6.

Table 4
Significantly Overrepresented Pathways for the Crohn’s Disease GWAS.

While genes that mediate immunological defects were expected to be among the most strongly associated with CD, surprisingly, a number of metabolic pathways, especially lipid metabolism, are enriched in the CD GWAS most strongly associated genes, all of which include the common pathway hits AMPK gamma and alpha (PRKAG2, PRKAA1) (discussed in detail in Supplemental Text). These metabolic defects suggest a basis for the genetic relationship between CD and both BD and CAD (Table 1) which is observed to be stronger than the correlation of CD with other autoimmune diseases (e.g., T1D).

Hypertension

Although a lack of genome-wide significant associations was observed for hypertension (HT) in the original WTCCC GWAS, a long list of significant and interconnected pathways was detected by our analysis of the most strongly associated SNPs from the GWAS of HT. The apparent high degree of interconnection between these pathways suggests that multiple, but related, mechanisms are likely to underlie hypertension susceptibility and, furthermore, demonstrates that a large number of common but low penetrance risk alleles are likely to be involved.

The most significant pathway overrepresented in the most strongly associated genes with HT suggests a role for defects in dopamine signaling. Specifically, genes that we among the most strongly associated with HT include the AMPA receptor (GRIA4), NMDA receptor (GRIN2A), GABA-A receptor (GABRB3) (receptors that play a role in the cytoskeletal remodeling of synapses in response to EphB signaling, which may be unrelated to hypertension), as well as ADCY5, GNAS, PPP3CA, and PRKAR2B. Dopamine signaling has a well established role in regulating blood pressure via a number of mechanisms, including regulation of vascular smooth muscle tone [21], which is another pathway significantly overrepresented in our analysis, and encompasses the receptor subunits CHRM2 and HTR2A, as well as ADCY4, ADCY5, GNAS, PRKAR2B, and CAMK2D.

These key members (ADCY4, ADCY5, GNAS, PRKAR2B, CAMK2D) also play a role in a wide variety of interconnected pathways including PKA signaling and cAMP signaling, as well as the development of alpha and beta adrenergic receptor signaling, which have well established roles in the control of blood pressure. Additionally, they are involved in pathways directly implicated in regulating blood pressure, including PGE2 signaling [22], and calcium signaling, with specific hits in AKAP6, calcineurin (PPP3CA) and its regulator calcipressin (DSCR1). Calcineurin inhibitors are known to increase blood pressure [23]. Calcineurin signaling, along with the previously mentioned key members, is also involved in the regulation of lipid metabolism, likely underlying the metabolic defects associated with hypertension risk. Glucose metabolism is also affected through ChREBP regulation with an additional hit in ACSL5. An unrelated pathway, plasmalogen biosynthesis, which is implicated in the control of lipoprotein levels, provides an additional genetic mechanism for metabolic defect in hypertension, with pathway members among the most significantly associated genes including PLD1, PLAG1B, and AGPS [24].

Finally, genetic defects in cell-cell interactions and cytoskeletal remodeling due to cell-cell signaling, as well as particular interactions between endothelial cells and smooth muscle cells, are implicated in this analysis, with associated genes Alpha-3/beta-1 integrin (ITAG3), Collagen IV (COL4A2), Claudin-14 (CLDN14), ACTN2, CTNNA3, and MAGI1. Genetic perturbations in these cell-cell interactions suggest a complimentary defect in the maintenance of vascular smooth muscle tone.

The evidence for the involvement of additional pathways in hypertension is sparse, but provides some interesting leads. Recent evidence, consistent with our analysis results, implicate chromosome condensation and apoptosis as potential pathogenic mechanisms [25]. Another interesting set of genes among the most strongly associated implicates macrophage inhibitory factor signaling, which include the associated genes mentioned above, as well as plasminogen PLA2G1B and CSNK2A2. A provocative connection between MIF and angiotensin II in spontaneously hypertensive rats has recently been observed [26]. Obviously, it will be interesting to see if further studies will verify the biological significance of these results.

The connection of MIF signaling and PGE2 processes involved in immune function may provide hints into the strong correlation between HT and CD and RA demonstrated in Table 1 [27]. Overrepresented biological processes for both diseases show great similarities and are presented in Supplemental Table 5.

Rheumatoid Arthritis

Rheumatoid arthritis (RA) is an autoimmune disease characterized by synovial inflammation and destruction of multiple joints. As would be expected, autoimmune diseases dominant the most overrepresented diseases among those associated with the most strongly associated RA genes, and immune functions dominant the most overrepresented processes (Supplemental Table 1, Supplemental Table 6). The pathway analysis implicates multiple specific pathways involved in the development of RA (Table 6). The top pathway implicates neuroendocrine defects, specifically in the secretion of macrophage migration inhibitory factor in the development of RA. Macrophage abundance significantly correlates with RA severity [28], and polymorphisms in the MIF promoter are significantly associated with early onset RA [28]. We find that genetic lesions affecting signaling proteins upstream of the release of MIF from pituitary cells, specifically GNG2, ADCY2, PRKCA, and ITPR3 are likely to increase MIF release, as elevated MIF levels are associated with RA [29]. Upon binding to receptors in macrophages, the PU.1 transcription factor, required for the expression of a variety of inflammatory factors is activated. Phosphorylation by Casein kinase 2 enhances the transactivation activity of PU.1 [30], and is identified as another associated gene in this pathway (CSNK2B). These four members, or a subset of these four members, GNG2, ADCY2, PRKCA, and ITPR3) are implicated in other significant pathways as well, suggesting a common genetic foundation for other signaling pathways with known roles in RA pathogenesis, such as prostaglandin E2 signaling [31].

Table 6
Significantly Overrepresented Pathways for the Rheumatoid Arthritis GWAS.

Expanding upon this genetic foundation, additional genes in calcium channels CACNG1, CAMK2D, and the MEF2B transcription factor further implicate IP3 signaling, cAMP signaling, G-protein signaling, and calcium signaling as general signaling pathways disturbed in RA. These pathways play roles in other well known RA-related biochemical disturbances such as modulation of immune functions in T lymphocytes and prostacyclin synthesis [32]. Additional associated genes implicate specific pathways which are likely to be differentially affected in different RA cases. For example, an additional gene in monocyte chemoattractant protein 1 (CCL2), implicated in leukocyte migration in the joint [33], suggests another mechanism through which the GNG2, ADCY2, PRKCA, and ITPR3 may be involved in development of RA. Associated genes for of MHC class II proteins, long known to underlie RA pathogenesis [34], with additional genes IL2RA, IL2RB and CD40, contribute to aberrant signaling between immune cells, and are clearly associated with RA in our study.

Another aspect of RA which shows a genetic component in our analysis is the expansion of neovascular networks into joints, and the role of these neovascular networks in bringing inflammatory cells into joints. Again, the factors discussed above play a significant role, with additional genes VCAM (VCAM1) and EGFR, indicate pathways known to be activated in RA, namely VEGF signaling [35], angiotensin signaling, [36] and crosstalk between the two pathways. Additional associated genes alpha-catenin (CTNNA3) and ZO-1 (TJP1), suggest genetic defects in cell barrier integrity contribute to vascular permeability and the infiltration of inflammatory cells into joints.

Other pathways implicated by our analysis, which may play a novel role in RA, but whose significance is unclear, are EDNRB signaling (hits: GUCY1A2, PRKG1) and PKA cytoskeletal reorganization (hits: AKAP13). Notably, EDNRB is downregulated in RA, providing some validation that this may be a true pathogenic pathway in RA [37]. Recently, AKAP13 was discovered as a regulatory of TLR2 signaling [38], and is likely to provide hints into the genetic connection between toll-like receptor signaling and RA [39].

Type 1 Diabetes

Type 1 Diabetes (T1D) is an early onset autoimmune disorder. As with RA, autoimmune diseases dominant the disease pathways, and immune functions dominate the processes, identified by our analysis of the most strongly associated genes with the T1D data (Supplemental Table 1 and Supplemental Table 7). Many genes previously described as having associations with T1D arose in our analysis including, MHC pathways, T lymphocyte pathways, interleukin 2 signaling pathways, and ERBB3 signaling pathways. Outside of these known single gene associations, we find additional associated genes CCR3, IL2, NFATC, TAP1 and TAP2 antigen peptide transporters, PSMD14 and PSMD9 proteasomes, AREG, NRG2, and NRG3 neuregulins (ERBB signaling mediators), all of which participate in pathways where the major players (i.e. the IL2 receptor, MHC proteins, and ERBB3 receptor) have already been associated with T1D. A large degree of overlap between T1D and RA GWAS results, as indicated by the high degree of correlation in Table 1, is also observed. Namely, overrepresented pathways involving IP3 receptors (ITPR1 and ITPR3), calcium channels (CACNA2D3), MEF2A, CAMK2D, and GNG2 suggest a similarity in the genetic perturbations underlying these diseases and, therefore, comparable immunological dysfunctions characteristic of the two conditions. Additional associated genes include STAT3, TNF, and a series of protein kinases (RAF1, MAPK14) which may explain some of the vascular complications associated with diabetes and JAK/STAT signaling augmented by angiotensin [40].

While immune pathways were to be expected to be overrepresented among T1D associated genes, an extremely provocative association with BRCA1 DNA repair was also observed and is discussed further in the Supplemental Text.

Type 2 Diabetes

Type 2 diabetes (T2D) is a metabolic disorder strongly associated with obesity, and is characterized by defects in glucose uptake in response to insulin. Our pathway analysis of T2D identified a small number of overrepresented pathways, although a number of interesting process defects were observed. The top significant pathway, regulation of lipid metabolism, has obvious implications for obesity. Gene members of this pathway among the most strongly associated with T2D include PPARG, which has already been implicated by multiple genome wide association studies. Downstream of PPARG, apolipoprotein C (APOC1), which is important for lipid metabolic processes, was found to be one of the most strongly associated genes with T2D. Upstream of PPARG we observe at the associated genes PLCB1, PRKCE, and PLA2G4A, all of which respond to nicotinic G-protein coupled receptors. Other associated genes in G-protein coupled receptor pathways, including BLNK, VAV2, and ARGHEF12, potentially alter signaling through numerous G-protein coupled receptor subtypes, of which at least types q and I signal through cdc42, another pathway hit. These parallel G-protein signaling pathways activate cdc42 and JNK resulting in insulin resistance [41]. Furthermore, these pathways are modulated by signaling through long chain fatty acids, such as arachidonic acid, as well as diacyl-glycerol, and with the addition of calcium signaling through CAMK2D, all converge upon AKT signaling, which is also implicated in our analysis, and plays a crucial role in promoting glucose uptake in response to insulin [42].

A connection between neurological diseases, especially Alzheimer disease, and type 2 diabetes, has been extensively catalogued in the literature [43]. Our disease pathway results further suggest that genes involved in neuronal disorders could also play a role in T2D (Supplemental Table 1), and neural development processes are overrepresented among the most strongly associated T2D genes (Supplemental Table 8). β-cells are known to share a large number of similarities with neuronal cells due to the absence of neuron-restrictive silencing factor in insulin producing cells [44]. TCF7L2 (TCF4), strongly associated with T2D, is also involved in the development of neuronal cells [45] and the pituitary gland [46]. Thus, our analysis suggests that β-cell developmental pathways, which share similarities to neuronal developmental signaling pathways (especially Wnt signaling), are likely to play a large role in T2D predisposition.

Discussion

Our analysis of the most strongly associated genes arising from the GWAS analysis of seven different diseases identified statistically significant evidence, along with biological plausibility, for pathogenic pathways involved in each. Some of these pathways are presumed to play a role in disease pathogenesis though variations in specific genes that have not yet been identified via standard single-locus-based associated analyses (for example: dopamine signaling in HT). These results strongly suggest that common human diseases are modulated by a large number of low risk genetic factors (‘polygenes’) which are not likely to be easily identified through the use of standard single locus-oriented, univariate GWAS analysis techniques. This is not counterintuitive with the data, considering that only a small number of genetic variations associated with disease at genome-wide significance levels (with relative risks greater than 1.3) have been identified by current GWAS initiatives, and the miniscule proportion of the disease burden explained by these identified variations. Pathway analysis is an extremely useful approach for extending current single locus-oriented, univariate GWAS analysis techniques since it seeks to extract large amounts of biologically relevant information from them. Although many variations in the genes we have identified in our analyses could possibly be identified in extremely large case/control GWAS initiatives, a pathway analysis of smaller studies is a much more efficient approach to identifying candidate risk factors, and may be a useful method for identifying specific genes for follow-up or replication studies. It should be noted that a “small GWAS,” is a relative term, corresponding to 2,000 cases per disease, which still translates into a significant financial investment that should be thoroughly leveraged to maximize its yield.

Our analyses strongly suggests that the pathogenic processes underlying susceptibility to any single common chronic are common across many diseases, and that the development of any particular disease phenotype is likely to be a function of multiple interacting general and specific risk factors. Most of the seven diseases in the WTCCC appear to be mediated by relatively weak risk factors in sets of general signaling strategies, including, G-protein, adenylate cyclase, protein kinase A and C, IP3, and calcium signaling mechanisms. These common risk factors are likely to underlie general morbidity and form the basis for the strong genetic similarity observed across what, on the surface, appear to be unrelated diseases. Built on top of these basic signaling risk factors are a wide variety of disease-specific risk factors, a large number of which are transcription factors, transcription regulatory factors, cell surface receptors, or other cell surface structures such as antigen presenting molecules. These risk factors are likely to be the driving force underlying the development of a specific disease phenotype over another.

Additionally, each disease appears to be mediated by at least two of three basic cellular functions: metabolic, neurological or inflammatory processes. These connections are best described in the context of surprising similarities among the diseases. One would not have anticipated that bipolar disorder would be demonstrated to have underlying pathways in common with CAD and T2D. While neurological functions were to be expected in the pathogenesis of BD, it is clear that lipid metabolism may also play an important role. HDL metabolism, specifically affecting ApoB, was predicted to contribute to BD. The high concentration of lipids within the nervous system, as well as the expanding number of associations implicating lipid metabolism deregulation to neurological disorders, such as ApoE mutations in Alzheimer’s disease, or α-synuclein mutations in Parkinson’s disease, confirm this connection [47]. While lipid metabolism has obvious implications in CAD and T2D, these diseases are also connected to BD through neurological defects. The neurological dysfunctions in CAD appear to stem from serotonin signaling and formation of neuromuscular junctions or the role of neuronal proteins after vascular injury. Both sources play an important regulatory role in the migration of VSMC, a central process in the pathogenesis of CAD [48,49]. The neurological connection in T2D appears to revolve around the development of β-cells and is exemplified by the strong association of T2D with TCF7L2 polymorphisms, a transcription factor recognized for its role in neural development before its association with T2D [64]. Another surprisingly strong connection involves HT with CD. In this case the connection of HT, considered a metabolic disease to CD, likely lies in the role of inflammation in vascular remodeling [50]. Similarly, metabolic disorders have a clear role in hypertension pathogenesis but also appear to play a role in CD. In CD, there are known disturbances in lipid regulation, which may contribute further to the inflammatory response and to development of metabolic syndrome in CD patients [51].

Similarly surprising connections between seemingly unrelated diseases have been observed previously, and many more will undoubtedly be revealed by upcoming GWAS. For example, myocardial infarction and macular degeneration, diseases whose phenotypes are seemingly unrelated, have been connected by susceptibility polymorphisms in CFH [52,53]. Myocardial infarction has also been connected to other inflammatory diseases, such as RA and multiple sclerosis, by MHC2TA polymorphisms [54]. Similarities between diseases are likely to reveal related pathological mechanisms, as has been observed in the case of Alzheimer’s disease and T2D, and may reveal therapeutic strategies effective across multiple diseases [55].

Our analysis is limited by the incompleteness of genomic coverage of the WTCCC and its lack of ability to discern structural variants and other human genomic variation beyond commonly occurring SNPs. However, the data and our pathway analysis and interpretation of it provide a new perspective on how large human genomic data set information can be processed to implicate the role of biologic pathways that would otherwise be left unrecognized and potentially show a convergence of disease mechanisms. Thus, beyond replication of particular SNP associations arising from future GWAS and sequencing studies, consideration and further validation of pathway analysis may be a useful and insightful research angle.

Methods

Basic Pathway Analyses

SNP data was downloaded from the WTCCC webpage (http://www.wtccc.org.uk/) [8]. SNPs were assigned to genes within 5 kB using Biomart (http://www.biomart.org/). SNPs mapping to multiple genes were assigned to a single gene according to the following hierarchy coding > intronic > 5’utr > 3’utr > 5’upstream > 3’upstream. The statistical evidence for the most significant SNP for each gene was assigned as a ‘weight’ for the corresponding gene in subsequent analyses. This strategy avoids issues with weight inflation induced by genes having different numbers of SNPs, multiple SNPs in LD within a gene, or deflation of the weight of the gene from multiple insignificant SNPs. 15,835 genes were assigned a weight in this manner. Correlations between the p-values of these genes were generated both by using Pearson’s correlation coefficients and Spearman’s correlation coefficients in JMP 5.11. Spearman’s correlations were conducted to demonstrate that outliers did not drive the Pearson correlation structure. Distances for the construction of a tree diagram were calculated as one minus the Pearson’s correlation coefficient computed over the –log p-values of the SNPs associated with each gene. Log p-values were used since they accentuate the weight difference for slightly significant vs. strongly significant genes. The tree diagram was generated using the neighbor joining method in MEGA 3.1 [56]. The top 2.5% of genes with corresponding Entrez ID’s were entered into MetaCore2 and tested for enrichment in Maps, Diseases, GO processes, and GeneGO processes, separately for each disease. Metacore uses a hypergeometric model to determine the significance of enrichment. A free trial of MetaCore is available at http://www.genego.com/productTrials.php.

Statistical and Biological Validation of Pathway Analyses Based on Ranking of Genes

To demonstrate that the pathways identified by our analyses are significant findings, rather than occurring by chance, the bottom or lowest ranking 2.5% of genes were subject to the same analysis as the top or highest 2.5% genes (Supplemental Table 9). In almost all cases, the pathways enriched in these analyses had no known or obvious biological association with the diseases in question. In fact, cell cycle and oxidative phosphorylation pathways are apparently unaffected across all diseases from our analyses, suggesting basic cellular processes are conserved in these disease states. However, the analysis of rheumatoid arthritis implicated a number of pathways involved in disease pathogenesis. The common thread connecting these immune system pathways, along with some of the non-immune system pathways, are hits in transcription factors RelA and c-Fos. The activities of these transcription factors are required for RA pathogenesis [57], their inhibition reduces inflammation [58,59], and their over expression promotes tumorogenicity [60,61]. Although these proteins are involved in RA pathogenic pathways, it is not surprising that these proteins are found among those genes not involved in RA pathogenesis. However, this indicates that care should be taken in the interpretation of pathway analysis, and deceptive false positive results are a possibility. On the other hand, the consistency and credibility of our results across multiple diseases indicates that such misleading false positive results are likely to be rare.

Statistical Validation of the Correlation Analysis Results Between Diseases

In order to demonstrate that the correlations between the diseases are not an artifact arising from the WTCCC’s use of common controls for all seven GWAS studies, we conducted a simulation study to determine the effect that the use of common controls would have on such correlations in an idealized case. The simulation was structured as follows: the number of genes and number of SNPs per gene were selected to match the WTCCC data. For each SNP a randomly determined allele frequency between 0.05 and 0.95 was selected. This frequency was used to randomly assign genotypes to 3,000 individuals per control group. Seven separate control groups were generated in this manner. Seven unique disease populations, composed of 2,000 members each, were generated by maintaining the same control genotype frequency per SNP within each disease population, with the addition of 5% probability per SNP that the genotype frequency would differ by ±5% relative to the controls. Therefore, minor allele variance at disease loci ranged from 5% vs. 10% (amounting to an odds ratio of 2.11) in cases vs. controls, to 45% vs. 50% in cases vs. controls (amounting to an odds ratio of 1.22). A p-value for each disease compared to each of the seven controls was then calculated. The negative log of the most significant SNP per gene was then assigned as the weight of that gene, as done in our WTCCC analyses.

The simulation allows for three types of comparisons: 1. the correlation between a single disease and its gene weights as determined by comparison to multiple control groups; 2. the correlation between different diseases and their gene weights as determined by comparison to identical controls; and 3. the correlation between different diseases and their gene weights as determined by comparison to different controls. 95% confidence intervals were then calculated for each set of correlations and the results are as follows: 1, the average correlation in GWAS results when the same case group is used but with different controls is r2 = 0.955 ± 0.0001 (n = 7350); 2. the average correlation between GWAS results when different case groups are used with the same control group is r2 = 0.391 ± 0.014 (n = 4200); and 3. the average correlation when different case groups with different control groups are used is r2 = 0.388 ± 0.014 (n = 47,250). In our idealized case it is clear that any baseline correlation between diseases is due to the similarity in non-disease gene genotype frequencies as determined by the increase in correlation between diseases using different control groups or the same control group (Δr2 ≈ 0.003), while a strong correlation is observed if the same case group is compared to multiple independent control groups. Due to population heterogeneity, baseline correlations in the WTCCC data can be expected to be, and are observed to be, much lower than those observed in our idealized simulation study, while some of the between-disease correlations in the WTCCC data approach the level of correlation observed between when identical case groups and different control groups, at least on the basis of our simulation studies. This lends strong evidence that our observed correlations are not mere artifacts, but rather reflect true GWAS results similarities between different disease and is further confirmed by the pathway similarity observed between these diseases.

To correct for p-value deflation due to mapping of multiple SNPs to a gene, we conducted a simulation study to determine the effect on the p-value resulting from selecting the lowest p-value from multiple SNPs, with equal frequency in case and control populations, mapping to a single gene in an idealized case. The simulation was structured as follows: 509 genes with 1 through 509 SNPs mapping to each gene were simulated. For each SNP a randomly determined allele frequency between 0.05 and 0.95 was selected. This frequency was used to randomly assign genotypes to 3,000 individuals per control group and 2,000 individuals per case group. A p-value for each SNP was then calculated, selecting the lowest p-value per gene. This process was repeated 500 times to determine the average p-value assigned to a gene with 1 through 509 SNPs mapping to it, and a correction factor was determined based upon the decrease in average p-value when more than one SNP mapped to a single gene.

To assess the overall similarity of the GWAS results across all the disease, one minus the correlation coefficient between the GWAS results for each disease was used as a measure of distance to create a tree diagram. This measure was chosen so that perfectly correlated data would correspond to no distance between diseases, increasing to a maximum distance of one between completely uncorrelated data. There appears to be three broad disease groupings as shown in Figure 1. The first group, BD, CAD and T2D, is composed of diseases with strong neurological and metabolic components. The second group, with RA and T1D, is composed of diseases which are primarily autoimmune disorders. Finally, the third group, with CD and HT, are intermediate disorders with inflammatory, neurological and metabolic aspects. This genetic categorization differs from commonly accepted categorizations, i.e. BD on its own; CAD, T2D and HT as metabolic disorders; and RA, T1D, and CD as autoimmune disorders. Thus, potential common genetically-mediated, pleiotropic, etiologic mechanisms may shed new light on disease classification.

Figure 1
Tree diagram of disease based upon the distances calculated as one minus the correlation coefficient between each disease. BD = bipolar disorder, CAD = coronary artery disease, T2D = type 2 diabetes, HT = hypertension, CD = Crohn’s disease, RA ...
Table 5
Significantly Overrepresented Pathways for the Hypertension GWAS.
Table 7
Significantly Overrepresented Pathways for the Type I Diabetes GWAS.
Table 8
Significantly Overrepresented Pathways for the Type II Diabetes GWAS.

Supplementary Material

Acknowledgments

NJS and his laboratory are supported in part by the following research grants: The National Heart Lung and Blood Institute Family Blood Pressure Program (FBPP; U01 HL064777-06); The National Institute on Aging Longevity Consortium (U19 AG023122-01); The NIMH-funded Genetic Association Information Network Study of Bipolar Disorder National (1 R01 MH078151-01A1); National Institutes of Health grants: N01 MH22005, U01 DA024417-01, and P50 MH081755-01; Scripps Genomic Medicine and the Scripps Translational Sciences Institute. AT is a Scripps Genomic Medicine Dickinson Scholar.

Footnotes

Publisher's Disclaimer: This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final citable form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.

1SAS Institute, Cary, NC, USA

2GeneGO Inc., Encinitas, CA

References

1. Bodmer W, Bonilla C. Common and rare variants in multifactorial susceptibility to common diseases. Nat Genet. 2008;40:695–701. [PMC free article] [PubMed]
2. Eberle MA, et al. Power to detect risk alleles using genome-wide tag SNP panels. PLoS Genet. 2007;3:1827–1837. [PMC free article] [PubMed]
3. Kotti S, Bickeboller H, Clerget-Darpoux F. Strategy for detecting susceptibility genes with weak or no marginal effect. Hum. Hered. 2007;63:85–92. [PubMed]
4. Loza MJ, et al. Assembly of inflammation-related genes for pathway-focused genetic analysis. PLoS ONE. 2007;2:e1035. [PMC free article] [PubMed]
5. Wang K, Li M, Bucan M. Pathway based approaches for analysis of genome-wide association studies. Am. J. Hum. Genet. 2007 In press. [PMC free article] [PubMed]
6. Wellcome Trust Case Control Consortium. Genome-wide association study of 14,000 cases of seven common diseases and 3,000 shared controls. Nature. 2007;447:661–678. [PMC free article] [PubMed]
7. Tsuang MT, Taylor L, Faraone SV. An overview of the genetics of psychotic mood disorders. J. Psychiatr. Res. 2004;38:3–15. [PubMed]
8. Niculescu AB, et al. Identifying a series of candidate genes for mania and psychosis: a convergent functional genomics approach. Physiol. Genomics. 2000;4:83–91. [PubMed]
9. Goossens D, et al. A novel CpG-associated brain-expressed candidate gene for chromosome 18q-linked bipolar disorder. Mol. Psychiatry. 2003;8:83–89. [PubMed]
10. Hallman DM, Srinivasan SR, Chen W, Boerwinkle E, Berenson GS. The beta(2)-adrenergic receptor Arg16-gly polymorphism and interactions involving beta(2)- and beta(3)-adrenergic receptor polymorphisms are associated with variations in longitudinal serum lipid profiles: the Bogalusa Heart Study. Metabolism. 2004;53:1184–1191. [PubMed]
11. Garcia-Portilla MP, et al. The prevalence of metabolic syndrome in patients with bipolar disorder. J. Affect. Disord. 2007 In press. [PubMed]
12. Owens GK, Kumar MS, Wamhoff BR. Molecular regulation of vascular smooth muscle cell differentiation in development and disease. Physiol. Rev. 2004;84:767–801. [PubMed]
13. Kleyn PW, et al. Identification and characterization of the mouse obesity gene tubby: a member of a novel gene family. Cell. 1996;85:281–290. [PubMed]
14. Kapeller R, et al. Tyrosine phosphorylation of tub and its association with Src homology 2 domain-containing proteins implicate tub in intracellular signaling by insulin. J. Biol. Chem. 1999;274:24980–24986. [PubMed]
15. Santagata S, et al. G-protein signaling through tubby proteins. Science. 2001;292:2041–2050. [PubMed]
16. McCaffery JM, et al. Common genetic vulnerability to depressive symptoms and coronary artery disease: a review and development of candidate genes related to inflammation and serotonin. Psychosom. Med. 2006;68:187–200. [PubMed]
17. Chung NA, et al. Angiogenesis, thrombogenesis, endothelial dysfunction and angiographic severity of coronary artery disease. Heart. 2003;89:411–415. [PMC free article] [PubMed]
18. Wang L, Fan C, Topol SE, Topol EJ, Wang Q. Mutation of MEF2A in an inherited disorder with features of coronary artery disease. Science. 2003;302:1578–1581. [PMC free article] [PubMed]
19. Im SH, Rao A. Activation and deactivation of gene expression by Ca2+/calcineurin-NFAT-mediated signaling. Mol. Cells. 2004;18:1–9. [PubMed]
20. Weaver JR, Good K, Walters RD, Kugel JF, Goodrich JA. Characterization of the sequence and architectural constraints of the regulatory and core regions of the human interleukin-2 promoter. Mol. Immunol. 2007;44:2813–2819. [PMC free article] [PubMed]
21. Zeng C, Zhang M, Asico LD, Eisner GM, Jose PA. The dopaminergic system in hypertension. Clin. Sci. 2007;112:583–597. [PubMed]
22. Yang T. Microsomal prostaglandin E synthase-1 and blood pressure regulation. Kidney Int. 2007;72:274–278. [PubMed]
23. Elliott WJ. Drug interactions and drugs that affect blood pressure. J. Clin. Hypertens. 2006;8:731–737. [PubMed]
24. Maeba R, et al. Plasmalogens in human serum positively correlate with high- density lipoprotein and decrease with aging. J. Atheroscler. Thromb. 2007;14:12–18. [PubMed]
25. Fuster JJ, Díez J, Andrés V. Telomere dysfunction in hypertension. J. Hypertens. 2007;25:2185–2192. [PubMed]
26. Sun C, et al. Lack of macrophage migration inhibitory factor regulation is linked to the increased chronotropic action of angiotensin II in SHR neurons. Hypertension. 2007;49:528–534. [PubMed]
27. Grundy SM. Inflammation, hypertension, and the metabolic syndrome. JAMA. 2003;290:3000–3002. [PubMed]
28. Kinne RW, Bräuer R, Stuhlmüller B, Palombo-Kinne E, Burmester GR. Macrophages in rheumatoid arthritis. Arthritis Res. 2000;2:189–202. [PMC free article] [PubMed]
29. Leech M, et al. Macrophage migration inhibitory factor in rheumatoid arthritis: evidence of proinflammatory function and regulation by glucocorticoids. Arthritis Rheum. 1999;42:1601–1608. [PubMed]
30. Lodie TA, et al. Stimulation of macrophages by lipopolysaccharide alters the phosphorylation state, conformation, and function of PU.1 via activation of casein kinase II. J. Immunol. 1997;158:1848–1856. [PubMed]
31. Robinson DR, Dayer J-M, Krane SM. Prostaglandins and their regulation in rheumatoid arthritis. Ann. N.Y. Acad. Sci. 1979;332:279–294. [PubMed]
32. Honda T, Segi-Nishida E, Miyachi Y, Narumiya S. Prostacyclin-IP signaling and prostaglandin E2-EP2/EP4 signaling both mediate joint inflammation in mouse collagen-induced arthritis. J. Exp. Med. 2006;203:325–335. [PMC free article] [PubMed]
33. Koch AE, et al. Enhanced production of monocyte chemoattractant protein-1 in rheumatoid arthritis. J. Clin. Invest. 1992;90:772–779. [PMC free article] [PubMed]
34. Gregersen PK, Silver J, Winchester RJ. The shared epitope hypothesis. An approach to understanding the molecular genetics of susceptibility to rheumatoid arthritis. Arthritis Rheum. 1987;30:1205–1213. [PubMed]
35. Carvalho JF, Blank M, Shoenfeld Y. Vascular endothelial growth factor (VEGF) in autoimmune diseases. J. Clin. Immunol. 2007;27:246–256. [PubMed]
36. Price A, et al. Angiotensin II type 1 receptor as a novel therapeutic target in rheumatoid arthritis: in vivo analyses in rodent models of arthritis and ex vivo analyses in human inflammatory synovitis. Arthritis Rheum. 2007;56:441–447. [PubMed]
37. Galligan CL, Baig E, Bykerk V, Keystone EC, Fish EN. Distinctive gene expression signatures in rheumatoid arthritis synovial tissue fibroblast cells: correlates with disease activity. Genes Immun. 2007;8:480–491. [PubMed]
38. Shibolet O, et al. AKAP13 a rhoa GTPase-specific guanine exchange factor is a novel regulator of TLR2 signaling. J. Biol. Chem. 2007;282:35308–35317. [PubMed]
39. Huang Q, Ma Y, Adebayo A, Pope RM. Increased macrophage activation mediated through toll-like receptors in rheumatoid arthritis. Arthritis Rheum. 2007;56:2192–2201. [PubMed]
40. Banes-Berceli AK, et al. Angiotensin II and endothelin-1 augment the vascular complications of diabetes via JAK2 activation. Am. J. Physiol. Heart Circ. Physiol. 2007;293:H1291–H1299. [PubMed]
41. Taniguchi CM, et al. The p85alpha regulatory subunit of phosphoinositide 3-kinase potentiates c-Jun N-terminal kinase-mediated insulin resistance. Mol. Cell. Biol. 2007;27:2830–2840. [PMC free article] [PubMed]
42. Manning BD, Cantley LC. AKT/PKB signaling: navigating downstream. Cell. 2007;129:1261–1274. [PMC free article] [PubMed]
43. Li L, Hölscher C. Common pathological processes in Alzheimer disease and type 2 diabetes: A review. Brain Res. Rev. 2007;56:384–408. [PubMed]
44. Atouf F, Czernichow P, Scharfmann R. Expression of neuronal traits in pancreatic beta cells. Implication of neuron-restrictive silencing factor/repressor element silencing transcription factor, a neuron-restrictive silencer. J. Biol. Chem. 1997;272:1929–1934. [PubMed]
45. Flora A, Garcia JJ, Thaller C, Zoghbi HY. The E-protein Tcf4 interacts with Math1 to regulate differentiation of a specific subset of neuronal progenitors. Proc. Natl. Acad. Sci. U S A. 2007;104:15382–15387. [PMC free article] [PubMed]
46. Brinkmeier ML, Potok MA, Davis SW, Camper SA. TCF4 deficiency expands ventral diencephalon signaling and increases induction of pituitary progenitors. Dev. Biol. 2007;311:396–407. [PMC free article] [PubMed]
47. Wenk MR. The emerging field of lipidomics. Nat. Rev. Drug Discov. 2005;4:594–610. [PubMed]
48. Vikenes K, Farstad M, Nordrehaug JE. Serotonin is associated with coronary artery disease and cardiac events. Circulation. 1999;100:483–489. [PubMed]
49. Clement CM, et al. Neuregulin-1 attenuates neointimal formation following vascular injury and inhibits the proliferation of vascular smooth muscle cells. J. Vasc. Res. 2007;44:303–312. [PubMed]
50. Intengan HD, Schiffrin EL. Vascular remodeling in hypertension: roles of apoptosis, inflammation, and fibrosis. Hypertension. 2001;38:581–587. [PubMed]
51. Ripollés-Piquer B, et al. Altered lipid, apolipoprotein, and lipoprotein profiles in inflammatory bowel disease: consequences on the cholesterol efflux capacity of serum using Fu5AH cell system. Metabolism. 2006;55:980–988. [PubMed]
52. Kardys I, et al. A common polymorphism in the complement factor H gene is associated with increased risk of myocardial infarction: the Rotterdam study. J. Am. Coll. Cardiol. 2006;47:1568–1575. [PubMed]
53. Hageman GS, et al. A common haplotype in the complement regulatory gene factor H (HF1/CFH) predisposes individuals to age-related macular degeneration. Proc Natl Acad Sci U S A. 2005;102:7227–7232. [PMC free article] [PubMed]
54. Swanberg M, et al. MHC2TA is associated with differential MHC molecule expression and susceptibility to rheumatoid arthritis, multiple sclerosis and myocardial infarction. Nat. Genet. 2005;37:486–494. [PubMed]
55. Ubeda M, Kemp DM, Habener JF. Glucose-induced expression of the cyclin-dependent protein kinase 5 activator p35 involved in Alzheimer's disease regulates insulin gene transcription in pancreatic beta-cells. Endocrinology. 2004;145:3023–3031. [PubMed]
56. Kumar S, Tamura K, Nei M. MEGA3: Integrated software for Molecular Evolutionary Genetics Analysis and sequence alignment. Brief. Bioinform. 2004;5:150–163. [PubMed]
57. Han Z, Boyle DL, Manning AM, Firestein GS. AP-1 and NF-kappaB regulation in rheumatoid arthritis and murine collagen-induced arthritis. Autoimmunity. 1998;28:197–208. [PubMed]
58. Lee UJ, et al. Dual knockdown of p65 and p50 subunits of NF-kappaB by siRNA inhibits the induction of inflammatory cytokines and significantly enhance apoptosis in human primary synoviocytes treated with tumor necrosis factor-alpha. Mol. Biol. Rep. 2007 In press. [PubMed]
59. Oliver SJ, et al. Vanadate, an inhibitor of stromelysin and collagenase expression, suppresses collagen induced arthritis. J. Rheumatol. 2007;34:1802–1809. [PubMed]
60. Sunters A, Thomas DP, Yeudall WA, Grigoriadis AE. Accelerated cell cycle progression in osteoblasts overexpressing the c-fos proto-oncogene: induction of cyclin A and enhanced CDK2 activity. J. Biol. Chem. 2004;279:9882–9891. [PubMed]
61. Chen X, Kandasamy K, Srivastava RK. Differential roles of RelA (p65) and c-Rel subunits of nuclear factor kappa B in tumor necrosis factor-related apoptosis-inducing ligand signaling. Cancer Res. 2003;63:1059–1066. [PubMed]
PubReader format: click here to try

Formats:

Related citations in PubMed

See reviews...See all...

Cited by other articles in PMC

See all...

Links

  • MedGen
    MedGen
    Related information in MedGen
  • PubMed
    PubMed
    PubMed citations for these articles