Using Fisher's method with PLINK ‘LD clumped’ output to compare SNP effects across Genome-wide Association Study (GWAS) datasets

Abstract

As the number of publically available GWAS datasets continues to grow, bioinformatic tools which enable routine manipulation of data are becoming increasingly useful. Meta-analysis using multiple GWAS datasets has become essential to elucidate novel SNP associations which may not be readily discovered in each GWAS individually due to insufficient power. Replication of GWAS findings is critical and is the ‘arbiter’ of genuine SNP associations. We have developed an ‘LD aware’ bioinformatics application which allows efficient comparison of SNP effects across multiple GWAS datasets using Fisher’s combined probability test from PLINK (v1.06) ‘LD clumped’ output. Availability: the application is freely available from the authors.

Motivation

Although genome wide association studies (GWAS) have given a valuable insight into the biological aetiology of many complex human diseases, insufficient power has meant a large number of these studies have failed to generate any significant ‘hits’ [1]. GWAS have often fallen below the numbers of cases and controls required to detect a modest effect (OR ∼1.25) from a common variant [1-3]. However, despite these power issues, they are still valuable for meta-analysis purposes. Combining individually underpowered GWAS will propel genuine associations and spurious associations will diminish [4].

Unfortunately, there are a number of constraints that have limited the effectiveness of whole-genome meta-analysis to date; i) given that GWAS use different genotyping platforms (Illumina or Affymetrix), each assaying different panels of SNPs, the number of ‘matched’ SNPs available for meta-analysis is limited. This is often confounded by SNP dropout during quality control procedures ii) whole genome meta-analysis is labour intensive without suitable bioinformatics software.

To overcome these constraints, we have developed an application written in PERL programming language which allows simple and efficient meta-analysis of multiple GWAS. Unlike other whole-genome meta-analysis tools, our application is ‘LD aware’ - if a single SNP is present in one study but is not present directly in the other, the p-value of the SNP is inferred using a perfect proxy (LD, r²=1).

We consider this application to have several advantages; i) maximize the number comparable SNPs (and therefore genome coverage) using LD (r²=1), ii) provide summary p-value statistics using Fisher's method, iii) annotate SNPs which have an odds ratio (OR) in opposing directions from different studies ('flippers'), and iv) tabulate the results to facilitate simple visualisation by conventional tools such as Microsoft Excel or SPSS

Implementation

General

PLINK (v1.06) [5] provides an ‘Id-clump’ analysis which allows automatic calculation of ‘clumps’ (blocks of SNPs in LD) across genotyp-ingchip platforms. The software we have developed uses the output file from this ‘LD clump’ analysis and performs the downstream analysis - meta-analysis of SNPs in each clump (taking one SNP/proxy (r²=1) from each study and combining their p-values).

The application consists of two files ‘perl.pl', ‘module.pm', where ‘perl.pl’ is an executable file when PERL language is installed. PERL programming language can be downloaded freely from http://www.perl.org/. ‘perl.pl’ file can be edited using a conventional text file editor, and allows users to define two parameters; i) the location and the filename of the input ('ld_clump') file generated in PLINK, and ii) the type of the study - case/control (CC) analysis or quantitative trait (QT) analysis. Fields requiring modification have been annotated in the ‘perl.pl’ file. The default input is case/ control analysis. Failure to adjust the parameter for the correct type of analysis will cause incorrect output in the results file.

To execute the programme, simply double click the file ‘perl.pl'; alternatively, type in ‘perl perl.pl’ into the command line. The command-line interface icon can be found in ‘Start -> All programs -> Accessories’ in Windows.

The programme was designed to handle an unlimited number of GWAS datasets and unlimited SNPs. However, currently we have only validated the application with a maximum of ten modelled datasets.

Issues and problem-solving

It is worth noticing that, when generating the ‘clump’ files, PLINK requires a reference GWAS dataset (usually HapMap) to calculate LD values between two SNPs. This data is freely available at http://hapmap.ncbi.nlm.nih.gov/. Given that LD values vary between ethnic populations, it is imperative that the reference dataset and the GWAS datasets are from a similar population thus avoidingstratification issues.

It is common that a SNP in one GWAS will have multiple perfect SNP proxies (r²= 1) in another independent GWAS. In this situation, although it is considered appropriate to use any pair of SNPs to perform meta-analysis, the application decides which proxy to use based on which proxy is closest to the index SNP. The distance between the SNP/proxy and the index SNP is annotated in the results file. It should be noted that if an index SNP is unique to one study and does not have a perfect proxy in any other studies, no meta-analysis results will be displayed for this SNP.

This application only uses perfect proxies (r² = 1). We acknowledge that this is a limitation as using imperfect proxies (r²< 1) will increase the number of comparable SNPs between studies. However, we do not recommend this as it will lead to inaccurate Fisher's combined probability test p-values. Currently, there is no weighting algorithm implemented therefore any SNP p-value inferred from a proxy with r²< 1 will be inaccurately treated as a perfect match. If the user wishes to use imperfect proxies, a minimal LD value of r²= 0.9 maybe a reasonable threshold. To use imperfect proxies, simply alter the (–clump-r2) parameter in PLINK, and run the application as usual. This may indeed be a valuable approach to increase coverage, analogous to imputation, but until the output is weighted accordingly the results will have to be interpreted with caution.

A ‘flipper’ refers to single SNP or SNP proxy in one dataset which has the opposite effect in the original study. Our application compares the OR (case/control analysis) or regression coefficient (BETA from quantitative trait analysis) of all SNP pairs from different studies, and annotates according to the following rules:

i) ‘YES - flipper’ - ORs are in opposite directions in two (or more) GWAS studies, irrespective of missing OR data in additional studies, ii) ‘NO - non-flipper’ - all SNPs OR in the same direction and OR data is present for all studies, and iii) ‘NA - not applicable’ - there is either no OR data, or datasets are missing OR data making it inappropriate to call a ‘non-flipper'. In studies with missing OR or BETA, the field has to be encoded as ‘-9’ in the ‘.assoc’ file for subsequent Id-clumping’ analysis in PLINK.

The application automatically recognizes the number of GWAS from PLINK Id-clump’ output files and tabulates the results accordingly. Although the application was designed for GWAS meta-analysis the user can perform analysis on much smaller datasets.

We suggest that this approach is used prior to more formal meta-analysis. It is essential that any potential finding that emerges using the application is verified by further investigation in a rigorous manner. Adjusting the genotypic data for covariates and taking into account heterogeneity between studies/samples will verify if observations involving both ‘flipping’ and ‘non-flipping’ alleles are likely to be genuine and worthy of downstream study.

EXAMPLE: Meta-analysis of Carrasquilloetal. 2009 [6], Reiman etal. 2007 [7] and Beecham et al. 2008 [8]GWAS

We used two Late-onset Alzheimer's disease (LOAD) GWAS datasets - Carrasquillo el al. 2009 and Reiman et al. 2007 and the ‘top hits’ tabulated in Beecham et al. 2008 to test the performance of the application. The sample sizes and genotyping platforms for the three studies are - Carrasquillo et al. 2009, 799 LOAD cases and 1199 controls on Illumina 300 chip, Reiman et al. 2007, 859 LOAD cases and 552 controls on Affymetrix 500K chip and Beecham et al. 2008, 492 LOAD cases and 496 controls on Illumina 550 chip. The total sample size of all three GWAS is 4,397 (2150 LOAD cases and 2247 controls).

No apolipoprotein E (APOE) related SNPs were listed in Beecham et al. 2008 ‘top hits'.

Procedure

Before using the software a number of PLINK analyses have to be undertaken:

1) Subject-level genotype data was obtained from Carrasquillo et al. 2009 and Reiman et al. 2007. The files were converted into PLINK format where necessary, and the SNP identifiers were converted into dbSNP ‘rs’ number.

2) A case/control allelic association test was undertaken using ‘-assoc’ using PLINK (v1.06) for Carrasquillo et al. 2009 and Reiman et al. 2007 GWAS datasets. This generates two '.assoc’ files ‘Carrasquillo.assoc’ and ‘Reiman.assoc'. Since we do not have the Beecham et al. 2008 GWAS dataset, a file called ‘Beecham.assoc’ was manually generated conforming to the format of an ‘.assoc’ file [8]. Three compulsory columns are required in the ‘.assoc’ file, and the headers have to be named exactly as ‘SNP', ‘OR’ and ‘P’ without quotation marks. All other information such as ‘BP', ‘CHR', ‘Al’ in the standard ‘.assoc’ file are not required, and can simply be ignored. As OR data was not included in the Beecham data, these values were set to ‘-9’ in the ‘OR’ column.

3) The Id-clump’ analysis was performed using PLINK (v1.06) with the three files we generated 'Carrasquillo. assoc', ‘Reiman.assoc’ and 'Beec/iam.assoc'. We downloaded the filtered version of HapMap data (CEU population, release 23) in PLINK binary format (.bed, .bim and .fam) from the PLINK website, http://pngu.mgh.harvard.edu/~purcell/plink/res.shtml’ [5]. The HapMap data downloaded contains 2.3 million SNPs.

The Id-clump’ analysis was performed using the following PLINK command:

• pink -bfile HapMapCEU23

• -clump Carras-quillo.assoc,Reiman.assoc,Beecham.assoc

• -clump-verbose

• -clump-annotate OR

• -clump-pl 1

• -clump-p2 1

• -clump-r2 0.99

• -out ld_clump

• -noweb

The PLINK method reads ‘-bfile’ (the HapMap data in PLINK format) and ‘clumps’ the three datasets based on HapMap LD r-squared values. ‘∼clump-r2 0.99’ ensures that only SNPs which are perfect proxies (r²>0.99) are clumped ('-clump-r2 1’ does not work). All SNPs were used to perform the Id-clump’ analysis irrespective of p-values ('-clump-p1 1’ and ‘-clump-p2 1'). The output of ORs was specified using ‘-clump-annotate OR’ (Use ‘-clump-annotate BETA’ for quantitative trait analysis). A single output file ‘ld_clump.clumped’ was generated using ‘-out ld_clump'. All these commands listed are compulsory to the subsequent analysis except ‘-clump-p1’ and ‘-clump-p2’ which allow the user to control p-value threshold.

4) The ‘perl.pl’ file was edited using a text file editor to ensure it contains the correct PLINK Id-clump’ output filename and correct type of analysis (either ‘CC for case/control analysis or ‘QT’ for quantitative trait analysis) as discussed earlier.

5) The application was executed by double clicking the ‘perl.pl’ icon in Windows. Results file named ‘results.txt was generated automatically.

Results

The top 7 results in Table 1 illustrate the utility of this application. The top 2 SNPs are in LD with the APOE locus and demonstrate highly significant p-values as expected (rs1114832 Fisher's method p-value 1.09 × 10^-9 and rs-10402271 Fisher's method p-value 1.13×10^-7); the first SNP met genome-wide significance (p = 1.67×10^-7) after correcting for the number of independent tests as described previously [4] and the second SNP approached this value.

Table 1.

Results from Meta-analysis of Carrasquillo et al. 2009, Reiman et al. 2007 and Beecham et al. 2008. The table shows Single Nucieotide Polymorphisms (SNPs) with Fisher's combined probability test p-value less than 1E-05. ‘Fl', ‘F2’ and ‘F3’ refers to the studies which have been inputted to perform the meta-analysis. ‘Fl’ indicated which study the ‘SNP’ is taken from (T, ‘2’ and ‘3’ refers to Carrasquillo et al. 2009, Reiman et al. 2007 and Beecham et al. 2008, respectively). ‘F2’ and ‘F3’ indicated which study ‘PR0XY1’ and ‘PR0XY2’ is taken from (T, ‘2’ and ‘3’ refers to the studies as above). ‘KB1’ and ‘RSQ1’ refer to the distance and LD between the index ‘SNP’ and ‘PR0XY1'. ‘KB2’ and ‘RSQ2’ refer to the distance and LD between the index ‘SNP’ and ‘PR0XY2'. The suffix (1,2 or 3) on column headers ‘Pvalue’ indicated p-value in each study individually as listed. ‘ClumpNo’ - index number ranked based on descending p-value in each study, ‘CHR’ - chromosome number, ‘FISHER’ - Fisher's combined probability test p-value and ‘FLIPPER’ - indicates whether the SNP is a ‘flipper’ (see main text for definition of a ‘flipper').

ClumpNo	SNP	F1	CHR	PROXY1	F2	KB1	RSQ1	PR0XY2	F3	KB2	RSQ2	FISHER	FLIPPER	Pvalue1	Pvalue2	Pvalue3
1	rslll4832	1	19	rslll4832	2	0	1	-	-	-	-	1.09E-09	NO	1.37 E-06	0.000799	-
4	rsl0402271	1	19	rsl0402271	2	0	1	-	-	-	-	1.13E-07	NO	4.54E-06	0.0248	-
2	rs2318144	1	8	rs6982990	2	0.532	1	-	-	-	-	3.57E-07	NO	2.22E-06	0.161	-
12	rs929156	3	6	rs929156	1	0	1	rs9261519	2	-22.6	1	4.32E-07	NA	0.0887	0.288	1.69E-05
20	rs3746319	3	19	rs3746319	1	0	1	rs2061332	2	1.43	1	4.34E-07	NA	0.985	0.0149	2.96E-05
7	rsll205641	3	1	rsll205641	1	0	1	rsll205641	2	0	1	1.10E-06	YES	0.385	0.34	8.41E-06
107	rs468345	2	21	rs468345	1	0	1	-	-	-	-	1.15E-06	NO	0.00353	0.000326	-
42	rs7679738	1	4	rs510115	2	2.67	1	-	-	-	-	1.63E-06	NO	9.72E-05	0.0168	-
5	rs9659092	3	1	rsl2022125	2	-83.5	1	-	-	-	-	1.83E-06	NA	-	0.404	4.54E-06
3	rs249153	2	12	rs249153	1	0	1	-	-	-	-	1.91E-06	NO	0.717	2.66E-06	-
68	rs9474661	2	6	rs4486000	1	-2.44	1	-	-	-	-	2.19E-06	NO	0.0142	0.000154	-
16	rs8039031	1	15	rs8039031	2	0	1	-	-	-	-	2.24E-06	NO	2.26E-05	0.0992	-
86	rs4693305	2	4	rs4693305	1	0	1	-	-	-	-	2.64E-06	NO	0.0119	0.000222	-
10	rs7318037	1	13	rs4456389	2	11.5	1	-	-	-	-	2.75E-06	YES	1.15E-05	0.239	-
6	rs3007421	1	1	rs3007421	2	0	1	-	-	-	-	3.06E-06	NO	6.54E-06	0.468	-
74	rs385771	1	5	rs385771	2	0	1	-	-	-	-	3.78E-06	YES	0.000163	0.0232	-
76	rsl0501120	1	11	rsl0501120	2	0	1	-	-	-	-	4.00E-06	NO	0.000171	0.0234	-
9	rs4313171	2	8	rs359819	1	-169	1	-	-	-	-	5.01E-06	NO	0.501	1.00E-05	-
144	rs6695249	1	1	rsl7113051	2	4.38	1	-	-	-	-	5.40E-06	NO	0.00045	0.012	-
11	rs3807031	3	6	rs3807031	1	0	1	-	-	-	-	5.73E-06	NA	0.494	-	1.16E-05
25	rs2119067	3	2	rs2119067	1	0	1	-	-	-	-	6.92E-06	NA	0.158	-	4.38E-05
35	rsllO33712	2	11	rsl2271660	1	26.9	1	-	-	-	-	7.85E-06	YES	0.127	6.18E-05	-
8	rs6546452	1	2	rsl7680828	2	9.37	1	-	-	-	-	8.28E-06	NO	8.55E-06	0.968	-
14	rs4759173	2	12	rsl0876820	1	-22.7	1	-	-	-	-	8.86E-06	NO	0.445	1.99E-05	-
22	rs2387100	3	13	rs2387100	1	0	1	rs9551404	2	-12.5	1	9.40E-06	NA	0.644	0.382	3.82E-05
31	rs7537266	2	1	rs7537266	1	0	1	-	-	-	-	9.52E-06	NO	0.186	5.12E-05	-
17	rsl3213247	2	6	rsl6892136	1	-115	1	-	-	-	-	9.55E-06	NO	0.417	2.29E-05	-
79	rs4904864	1	14	rsl0484035	2	14	1	-	-	-	-	9.58E-06	YES	0.000186	0.0515	-

Sub-significant hits may prove to be genuine when more datasets are included. rs2318144 is located 200kb upstream of the inositol mono-phosphatase domain containing 1 gene, IMPAD1; as of October 2010 this gene has yet to figure as an AD candidate in the AlzGene forum [3].

rs929156 (p = 4.32×10^-7) is located in an exon of tripartite motif containing protein 15 (TRIM15) gene. There is evidence that TRIM15 is involved in the innate immune system however, its association with Alzheimer's disease has yet to be sufficiently investigated [9-11].

rs3746319 (p = 4.34×10^-7) was found to be located in an exon of the zinc finger protein (ZNF224). Although this SNP is close to the APOE region, the effect has been suggested to be independent of APOE status [8].

rsll205641 demonstrates a dichotomy of effect i.e. it's OR is not compatible between data-sets (shows opposing effects) and is thus indicated as a ‘flipper'.

rs468345 (p = 1.15×10⁶) is located ∼120kb upstream of Amyloid-beta precursor protein (APP) gene. It is well known that APP proteolysis generates beta amyloid (A(3), and extracellular deposition of Ap plaques is a hallmark of Alzheimer's disease [12]. Ap is the central player in amyloid cascade hypothesis in AD [11, 13].