Linkage-Disequilibrium-Based Binning Affects the Interpretation of GWASs

Abstract

Genome-wide association studies (GWASs) are critically dependent on detailed knowledge of the pattern of linkage disequilibrium (LD) in the human genome. GWASs generate lists of variants, usually SNPs, ranked according to the significance of their association to a trait. Downstream analyses generally focus on the gene or genes that are physically closest to these SNPs and ignore their LD profile with other SNPs. We have developed a flexible R package (LDsnpR) that efficiently assigns SNPs to genes on the basis of both their physical position and their pairwise LD with other SNPs. We used the positional-binning and LD-based-binning approaches to investigate whether including these “LD-based” SNPs would affect the interpretation of three published GWASs on bipolar affective disorder (BP) and of the imputed versions of two of these GWASs. We show how including LD can be important for interpreting and comparing GWASs. In the published, unimputed GWASs, LD-based binning effectively “recovered” 6.1%–8.3% of Ensembl-defined genes. It altered the ranks of the genes and resulted in nonnegligible differences between the lists of the top 2,000 genes emerging from the two binning approaches. It also improved the overall gene-based concordance between independent BP studies. In the imputed datasets, although the increases in coverage (>0.4%) and rank changes were more modest, even greater concordance between the studies was observed, attesting to the potential of LD-based binning on imputed data as well. Thus, ignoring LD can result in the misinterpretation of the GWAS findings and have an impact on subsequent genetic and functional studies.

Main Text

Over the past decade, genome-wide association studies (GWASs) have revolutionized the analysis of human complex genetic traits. By scanning hundreds of thousands of genetic variants, typically SNPs, in hundreds or thousands of individuals, they search for the variant(s) that associate with a particular disease or trait. Critical to the development and evolution of GWASs has been the creation of the International HapMap Project,¹ which has cataloged the common patterns of human genetic variation, including the linkage disequilibrium (LD) between SNPs. Knowledge of this LD, or nonrandom association of alleles at multiple loci, has made it possible to identify informative subsets of SNPs (i.e., “tagging SNPs”) that capture the bulk of genome-wide variation and has resulted in affordable genome-wide genotyping. To date, almost 1,000 GWASs have been published and have tested hundreds of human traits and reported thousands of significant associations (Catalog of Published Genome-Wide Association Studies²). Previously known associations have been confirmed, and new candidates have been implicated.³ However, a general sense of disappointment lingers because GWASs have fallen short of the initial expectation that they would unravel the genetic basis of complex traits.^4,5 Recent analyses reveal that a large proportion of the “missing heritability”^5,6 can be explained by a polygenic model that considers all GWAS SNPs simultaneously,^7–9 but these studies provide no clues about the identity of the susceptibility variants or the underlying biology of the trait.⁶ Thus, much attention has been given to uncovering and characterizing this “missing” or “hidden” heritability.^6,10

In a conventional GWAS, each SNP is considered separately (the “single-marker” approach), resulting in a list of variants ranked according to the statistical significance of their association to the trait (i.e., their p value).¹¹ The “top hits” are typically reported, and the relevance of each finding, as well as the focus of future work, is primarily based on the functional unit(s), namely gene(s), implicated by the associated SNP. Furthermore, gene-based methods are increasingly being applied as complementary approaches to the analysis of GWAS data. These methods take the gene instead of the individual SNP as the basic unit of association and thus allow aggregation of SNPs of smaller effect, potentially increasing power and reducing the multiple-testing burden.^12–14 They enable the incorporation of biological knowledge for greater insight into the mechanisms underlying the trait and are essential for subsequent pathway-based approaches.¹³ Gene-based methods also facilitate direct comparison of independent studies because they are unaffected by allelic heterogeneity and potential differences in SNP coverage and LD patterns.¹⁵

The success of both single-marker and gene-based approaches is critically dependent on the correct assignment of SNPs to genes. At the single-marker level, the aim is to identify the gene(s) that the associated SNP is tagging. At the gene level, the aim is to attribute all SNPs tagging a particular gene to that gene. Although LD can span hundreds of kilobases,^16,17 when GWAS results emerge, the SNPs of interest are typically assigned to the nearest gene or transcript within a specified distance.¹⁴ In turn, genes are typically represented only by the SNPs that are physically located within the transcribed region or predefined flanking region.¹³ It is not systematically taken into consideration that an associated SNP might be in high LD with another SNP (genotyped or not) located hundreds of kilobases away in a different gene or that a genotyped SNP positioned outside the defined boundaries of a gene is tagging that gene. Here, we show that ignoring LD discards valuable information and potentially leads to the incorrect localization of the association signal and might mislead the interpretation of GWAS data.

We have therefore developed a flexible R package (LDsnpR) that systematically assigns SNPs to genes (or relevant predefined genome “bins”) by using SNP association results (e.g., p values), bin definitions, and precalculated pairwise LD data (e.g., r² values) provided by the user (Figure S1, available online). By default, LDsnpR assigns a SNP to a bin if that SNP is located within the physical boundaries of that bin (i.e., the “positional-binning” approach). Then, as a unique feature of this package, the user has the option of also assigning a genotyped SNP to a bin if that SNP is in high pairwise LD with another SNP (genotyped or not) located within the physical boundaries of that bin (i.e., “LD-based-binning” approach). Although a genotyped SNP cannot be assigned to a particular gene more than once, it can be assigned to more than one gene.

As proof of principal, we used LDsnpR to assess the impact of the LD-based-binning approach (versus the positional-binning approach) on the results of three published GWASs on bipolar disorder (BP), each unimputed and genotyped on a different platform. The three GWASs are (1) the UK-based Wellcome Trust Case Control Consortium (WTCCC) BP GWAS,¹⁸ (2) the Norwegian Thematically Organized Psychosis (TOP) BP GWAS,¹⁹ and (3) a German BP GWAS²⁰ (Table 1). Each GWAS had been previously approved by the relevant local research ethics committees, and all participants had provided written informed consent.^18–20 In addition, we assessed the impact of LD-based binning on imputed versions of the TOP and German GWASs, in which ungenotyped markers had been statistically inferred¹¹ on the basis of LD from different reference panels (i.e., HapMap Phase III for TOP; HapMap Phase III and 1,000 Genomes²¹ for German) (Table 2).

Table 1.

Study Descriptions and Summary of Coverage for Positional-Binning and LD-Based-Binning Approaches for Original, Unimputed Datasets

	WTCCCa			TOPb			Germanc
Sample size (cases/controls)	1,868/2,938			198/336			682/1,300
Platform used	Affymetrix 500K			Affymetrix6.0			Illumina HumanHap550v3
Number of post-QC SNPs for binning	468,648			615,396			511,978
Binning data	Positional binning	LD-based binning	Differenced	Positional binning	LD-based binning	Differenced	Positional binning	LD-based binning	Differenced
Number of genes coverede	30,610 (83.4%)	33,443 (91.1%)	2,833 (9.3%)	31,823 (86.7%)	33,905 (92.4%)	2,082 (6.5%)	31,708 (86.4%)	33,861 (92.3%)	2,153 (6.8%)
Number of post-QC SNPs binned	237,869 (50.8%)	277,534 (59.2%)	39,665 (16.7%)	307,949 (50.0%)	363,570 (59.1%)	55,621 (18.1%)	272,914 (53.3%)	308,634 (60.2%)	35,720 (13.1%)
Number of SNPs binned to only 1 gene	199,752 (84.0%)	178,544 (64.3%)	21,208 (10.6%)	259,223 (84.2%)	234,036 (64.4%)	25,187 (9.7%)	228,098 (83.6%)	209,458 (67.9%)	18,640 (8.2%)
Number of SNPs binned to ten or more	135 (0.057%)	2,537 (0.91%)	2,402	174 (0.057%)	3,106 (0.85%)	2,932	141 (0.052%)	2,072 (0.67%)	1,931
Mean number of SNPs per bin (median)	9.4 (4)	15.2 (10)	6.6 (4)	11.7 (5)	19.4 (13)	8.4 (6)	10.5 (5)	15.4 (10)	5.6 (4)
Range (min–max)	1–514	1–515	0–87	1–687	1–701	0–112	1–655	1–665	0–64
Number of genes with only one SNP	4,830 (15.8%)	1,531 (4.6%)	3,299 (68.3%)	3,604 (11.3%)	992 (2.9%)	2,612 (72.5%)	3,647 (11.5%)	595 (1.8%)	3,052 (83.7%)

The following abbreviation is used: QC, quality control.

^aThe UK-based Wellcome Trust Case Control Consortium (WTCCC) BP GWAS.¹⁷

^bThe Norwegian Thematically Organized Psychosis (TOP) BP GWAS.¹⁸

^cA German BP GWAS.¹⁹

^dPercentages indicate percent increase or decrease from positional to LD-based binning.

^eEnsembl 54 (May 2009) genes (total N = 36,693) tagged by at least one SNP.

Table 2.

Study Descriptions and Summary of Coverage for Positional-Binning and LD-Based-Binning Approaches for Imputed Datasets

	TOPaImputedb			GermancImputedb
Sample size (cases/controls)	198/336			657/1,308
Imputation reference panel	HapMap Phase III (CEU)			1,000 Genomes (pilot 1, CEU) and HapMap Phase III (CEU)
Post-QC SNPs for binning	992,161			4,825,148
Binning data	Positional binning	LD-based binning	Differenced	Positional binning	LD-based binning	Differenced
Number of genes coverede	33,242 (90.6%)	34,193 (93.2%)	951 (2.9%)	32,116 (87.5%)	32,259 (87.9%)	143 (0.4%)
Number of post-QC SNPs binned	521,720 (52.6%)	612,316 (61.7%)	90,596 (17.4%)	2,394,441 (49.6%)	2,613,493 (54.2%)	219,052 (9.1%)
Number of SNPs binned to only one gene	431,808 (43.5%)	367,671 (37.1%)	64,137 (14.9%)	1,979,660 (41.0%)	1,855,413 (38.5%)	124,247 (6.3%)
Number of SNPs binned to ten or more	267 (0.03%)	7,967 (0.8%)	7,700	1,272 (0.03%)	16,807 (0.3%)	15,535
Mean number of SNPs per bin (median)	19.3 (9)	35.9 (25)	17.1 (12)	91.6 (44)	130.6 (84)	39.5 (26)
Range (min–max)	1–1,046	1–1,062	0–214	1–5,570	1–5,573	0–573
Number of genes with only one SNP	1,795 (5.4%)	651 (1.9%)	1,144 (63.7%)	241 (0.8%)	208 (0.6%)	33 (13.7%)

The following abbreviation is used: QC, quality control.

^aThe Norwegian Thematically Organized Psychosis (TOP) BP GWAS.¹⁸

^bImputation details: the Norwegian TOP dataset was imputed according to the ENIGMA protocol with the use of MACH imputation software³⁸ and HapMap Phase III (CEU) as the reference panel. The German dataset was imputed with IMPUTE2 software³⁹ and the 1,000 Genomes Project (Pilot 1, CEU) and HapMap Phase III (CEU) as reference panels.

^cA German BP GWAS.¹⁹

^dPercentages indicate percent increase or decrease from positional to LD-based binning.

^eEnsembl 54 (May 2009) genes (total N = 36,693) tagged by at least one SNP.

BP is a severe complex psychiatric disorder that shows high heritability (60%–80%) but for which clear genetic risk factors remain elusive.⁴ Although several GWASs on BP have been performed (Catalog of Published Genome-Wide Association Studies²), the findings have shown little overlap at both the SNP and gene levels. Also, only a handful of SNPs have achieved genome-wide significance (<∼10⁻⁸), and these SNPs only explain less than 3% of the heritability,^4,22 suggesting that psychiatric disorders, such as BP, might be less amenable to GWASs than other disorders.^5,23 However, systematic LD-based gene binning has not been applied to these datasets, possibly contributing to the apparent lack of success. Thus, we assessed the effects of the LD-based-binning approach relative to the traditional positional-binning approach with respect to (1) gene coverage, (2) changes in the results and, potentially, the interpretation of findings, and (3) pairwise concordance of the findings among the BP GWASs.

In brief, for LDsnpR, gene bin definitions were based on the Human Ensembl release 54 (May 2009) gene identifiers with unambiguous positional information (N = 36,693). We extended these gene bins by another 10 kb on either side to best capture potential regulatory regions.^24,25 The LD data were based on HapMap Phase II release 27 and were restricted to that of the CEU (Utah residents with ancestry from northern and western Europe from the CEPH collection) sample. We set the pairwise LD at the widely accepted threshold of r² ≥ 0.8²⁶ to limit the loss of power needed for the detection of association at the linked locus.²⁷

We first compared the extent of coverage between the positional-binning and LD-based-binning approaches in the published, unimputed datasets (Table 1). By allowing us to identify the intergenic SNPs that tag genes, LD-based binning resulted in a ∼13%–18% increase in the number of SNPs included in the gene-binning process. Intergenic SNPs represent ∼40% of GWAS trait-associated SNPs.³ Notably, LD-based binning “recovered” >2,000 genes (>6%) in all three datasets, increasing the proportion of Ensembl 54 genes tagged by at least one SNP from ∼83% to >91%. Furthermore, there was an increase in the density of coverage; an average of 5.6 to 8.4 (median of four to six) SNPs were added per gene, and there was an overall decrease (>68%) in the number of genes tagged by only one SNP.

The imputed datasets also yielded increased coverage (Table 2) but, as expected, to a lesser extent depending on the reference panel used for imputation. Although HapMap II (i.e., LDsnpR reference panel) is denser than HapMap III²⁸ (i.e., reference panel for the TOP and German studies), imputation on the 1,000 Genomes data (i.e., reference panel for the German study) potentially gives the densest coverage. For the TOP and German imputed datasets, LD-based binning resulted in an increase of 17.4% and 9.1%, respectively, in the number of SNPs included in the gene-binning process and the recovery of 951 (2.9%) and 143 (0.4%) genes, respectively. Although this is only a small proportion of the total gene coverage, the recovery of these genes enables them to be considered as candidates for BP association and might lead to a better understanding of the biology should the true association stem from them. Also of note, in the German GWAS, LD-based binning alone achieved an overall gene coverage of 92.3% (imputation achieved 87.5% coverage, and imputation combined with LD-based binning achieved 87.9% coverage), suggesting that under some scenarios, LD-based binning alone can offer the most coverage. As with the original GWASs, there was an increase in the density of coverage; an average of 17.1 and 39.5 (median 12 and 26) SNPs were added per gene for the TOP and German imputed datasets, respectively. There was also a decrease in the number of genes tagged by only one SNP (63.7%); the decrease was not as notable for the German imputed dataset (13.7%).

We next assessed the effects of the LD-based-binning approach on the results of the three GWASs at both the single-marker and gene levels. At the single-marker level, we used the positional-binning and LD-based-binning approaches to compare the genes tagged by the most significant SNPs reported in the original publications^18–20 (Table S1). Although LD-based binning made no difference to the results of the TOP BP study, three of the 14 SNPs in the WTCCC BP study and three of the eight SNPs in the German BP study implicated additional or alternative genes. Interpreting GWAS single-marker results demands fastidious consideration because when given only the p value, it is not immediately clear where the true source of the association originates¹⁷ and thus which is the true candidate gene. The overall potential for mislocalizing the association signal was underscored by the reduced number of SNPs tagging only one gene and the increased number of SNPs tagging ten or more genes after LD-based binning (Tables 1 and 2). Further investigations, such as expression studies,²⁰ are therefore warranted before attributing putative causality to a gene and, as a result, nominating it as the focus of future fine-mapping, functional, and other expensive and time-consuming follow-up studies.²⁹

As previously stated, gene-based analyses are ideal for pathway approaches, which aid in the interpretation of GWAS results by exploiting prior biological annotation to determine whether certain biological functions are enriched (i.e., overrepresented) among the more significant genes in a dataset. These methods require one measure of association (or score) for each gene on the basis of the individual SNP association signals. Here, we used a function in LDsnpR to score each gene with the most significant p value (i.e., the minimum p value approach), which was adjusted for the number of SNPs tagging that gene by a modification of Sidak's correction.³⁰ The minimum p value approach is the most widely used gene-scoring approach³¹ and assumes an underlying genetic architecture in which a single SNP, or locus, within the gene contributes to the disorder. The modification performs at least as well as a powerful regression-based method in correcting for the bias due to SNP number.³² In this study, the correlation between the gene score and the number of SNPs in the bin was reduced from Pearson r² > 0.30 to r² < 0.020 in all three datasets after the modified Sidak correction was applied. Also, permutation-based gene-set analysis, as implemented in PLINK,³³ on the German GWAS confirmed the high correlation between modified Sidak-corrected p values and permutation-based p values (r² > 0.95). The genes were scored for both the positional-binning and LD-based-binning approaches and were compared.

The overall correlation in the ranks of the genes between the two approaches was <0.83 in the three original datasets and the TOP imputed dataset, indicating that LD-based binning altered the scores and the subsequent ranks of the genes. Although not as large, changes in rank were also observed in the German imputed dataset (Table 3). When a resampling analysis was performed on the unimputed WTCCC dataset (it randomly excluded 5% of the samples [20 repetitions]), the average overall correlation in ranks due to LD-based binning (0.80) was lower than that resulting from random fluctuations in the datasets (>0.87), indicating greater changes due to LD (Table S2). Such changes in rank are likely to impact threshold-free, rank-based pathway approaches, such as gene-set-enrichment analysis,³⁴ which aims to determine whether a predefined set of genes is enriched at the top of a ranked list. By inspecting the top 2,000 genes emerging from the two binning approaches, we found a 27%–34% difference between the two gene lists in the three unimputed and the TOP imputed datasets and a 15.5% difference in the German imputed dataset. Here, the resampling analysis in the WTCCC GWAS found that random fluctuations in the dataset led to a 25.6% change in the top 2,000 genes, whereas LD-based binning resulted in a 30.7% difference (Table S2). For threshold-based approaches, such as Ingenuity Pathway Analysis and ALIGATOR,³⁵ in which a list of genes meeting a specified threshold is tested for overrepresentation of a particular biological function, LD-based binning could result in the submission of a substantially different list. Changes in the ranks of the genes are thus likely to impact the outcome of these analyses and possibly the overall biological interpretation of the findings. The extent to which these LD-based changes are meaningful will also depend on the study design and resulting power, given that the resampling analysis shows that substantial changes in results can also occur as a result of slight changes in the dataset.

Table 3.

Effect of LD-Based Binning on Ranks of Genes within Each GWAS

	WTCCC	TOP	German	TOP Imputed	German Imputed
Correlationa of gene ranks	0.79	0.83	0.83	0.83	0.92
Number of genes moving into top 2,000 with LD-based binning	681 (34.0%)	601 (30.0%)	538 (26.9%)	558 (27.9%)	309 (15.5%)

^aSpearman rank correlation (i.e., rho).

Finally, we assessed whether LD-based binning improved the concordance of results across studies, especially in light of the aforementioned changes in the ranks of the genes. We compared the positional-binning and LD-based-binning approaches by performing pairwise rank-correlation analyses of the three GWAS datasets at both the SNP level and the gene level (Table 4). When the positional-binning approach was used, little to no correlation was observed at both the SNP and gene levels. However, with LD-based binning, the overall rank correlation increased by ∼3% and was more significant for all pairwise comparisons, including the imputed datasets. Interestingly, the greatest concordance was observed when LD-based binning was combined with imputation, highlighting the complementary nature of the two methods. Although there was no obvious increase in overlap in the top gene hits (data not shown), this increase in overall concordance warrants the use of the LD-based-binning approach for the reanalysis of these and other datasets in the search for common functional gene sets and pathways. The observed increase in correlation persisted even when regions of high LD, such as the MHC (major histocompatibility complex) region on chromosome 6, were excluded (data not shown).

Table 4.

Pairwise Concordance between GWASs at SNP and Gene Levels

	WTCCC vs. TOP	WTCCC vs. German	TOP vs. German	TOP Imputed vs. German Imputed
SNP level	0.0066 (0.00018)	0.0037 (0.31)	−0.0018 (0.51)	−0.00023 (0.83)
Gene level (positional binning)	0.030 (1.78 × 10−7)	−0.0017 (0.78)	0.023 (4.78 × 10−5)	0.068 (<2.2 × 10−16)
Gene level (LD-based binning)	0.077 (<2.2 × 10−16)	0.027 (7.24 × 10−7)	0.053 (<2.2 × 10−16)	0.098 (<2.2 × 10−16)

The Spearman rank correlation and p value (in parentheses) are shown for each pairwise comparison.

Our study illustrates the importance of systematically accounting for LD in the interpretation of GWAS results. To the best of our knowledge, our study is the first to quantify the added value of LD-based binning; in particular, it shows an increase in the concordance of results across independent GWASs of a trait as complex as BP. Excluding LD defies the basic premise of the GWAS approach by discarding valuable genetic information and risking the incorrect localization of the association signal and the misinterpretation of the biology of the findings. Our findings call for a reanalysis of previously published GWAS data via the LD-based-binning approach and for future GWASs to adopt this method automatically. LDsnpR facilitates this process by efficiently assigning SNPs to genes and provides the option of scoring the genes for direct entry into pathway-analysis tools. LDsnpR's flexible framework allows the application of different gene-scoring methods; the application of such methods is necessary for detecting gene-based associations under different genetic architectures for the traits.³¹ The user-definable r² parameter enables the scanning of a greater range of allele frequencies at the linked locus.²⁷ Bin definitions and precalculated pairwise LD information can be updated on the basis of the user's interests and the information available. LD-based binning might also serve as a complementary and/or alternative approach to imputation. In particular, as high-quality LD data from the 1,000 Genomes Project²¹ emerges, all GWASs, including those previously subjected to imputation, might benefit from simple and efficient LD-based binning at no extra cost. As we show here, LD-based binning can further enhance imputed GWASs, albeit to a lesser extent than unimputed datasets. More tools that allow for incorporation of LD into the interpretation of GWAS data are emerging,^36,37 further testifying to the importance of this approach. Also, for studies genotyped on different platforms and/or imputed with the use of different reference panels, LD-based binning enables uniform comparison at both the gene and pathway levels.

It is crucial to note that our study, as well as LDsnpR, only addresses SNP-to-gene assignment. Issues involving the derivation of the most accurate gene score (which accounts for gene size and LD between SNPs), the handling of SNPs that are assigned to multiple, possibly overlapping, genes, and the correlation between genes are unresolved obstacles for pathway-analysis approaches¹³ and are beyond the scope of this paper. Furthermore, the benefits of LD-based binning will be unique to each GWAS depending on the trait and its true underlying genetic architecture, the study design, and the extent of SNP coverage.

Web Resources

The URLs for data presented herein are as follows:

• 1,000 Genomes Project, http://www.1000genomes.org/

• Catalog of Published Genome-Wide Association Studies, www.genome.gov/gwastudies/

• ENIGMA protocol, http://enigma.loni.ucla.edu/protocols/genetics-protocols/

• HapMap Project, http://hapmap.ncbi.nlm.nih.gov/

• Human Ensembl Release 54, http://may2009.archive.ensembl.org/biomart/martview/11839bb5ec82fb10bf0333540fa09c46

• IMPUTE2 Software, http://mathgen.stats.ox.ac.uk/impute/impute_v2.html

• Ingenuity Pathway Analysis, http://www.ingenuity.com/

• LDsnpR, http://services.cbu.uib.no/software/ldsnpr

• PLINK, http://pngu.mgh.harvard.edu/∼purcell/plink/

• R Archive Network, http://cran.r-project.org