Gene-based tests are now well established as complementary methods to traditional per-single nucleotide polymorphism (SNP) GWAS. These methods test for enrichment of multiple SNPs associated with the disease/trait that individually have too modest an effect on the phenotype to reach genome-wide significance using a per-SNP test. A key issue is accounting for linkage disequilibrium (LD) and gene size (number of SNPs). A permutation approach where phenotype labels are shuffled while keeping the markers fixed is considered the gold standard for correcting for LD and SNP number. However, this approach is computationally intensive and can only be applied to GWASs on unrelated individuals. We have previously shown a simulation approach generates similar results to the permutation (Liu et al., Reference Liu, McRae, Nyholt, Medland, Wray, Brown and Macgregor2010). The VEGAS approach is computationally tractable and can be applied to any GWAS experimental design (unrelated individuals, family designs, DNA pooling designs). Novel loci not identified using per SNP tests have been found using VEGAS (Cheng et al., Reference Cheng, Schache, Ikram, Young, Guggenheim, Vitart and Baird2013). Imputation to the HapMap reference panel has been superseded by the availability of the 1,000 Genomes phase 1 data (around 38 million variants; Genomes Project et al., Reference Genomes Project, Abecasis, Auton, Brooks, DePristo, Durbin and McVean2012). By updating VEGAS to use 1,000 Genomes phase 1 data, we are able to improve our LD estimates given the increase in the size of the reference panel (e.g., N for European ancestry subset is 379 compared to 90 in HapMAP phase 2), as well as updating genome build from hg18 to hg19 (Genomes Project et al., Reference Genomes Project, Abecasis, Auton, Brooks, DePristo, Durbin and McVean2012; International HapMap et al., Reference International HapMap, Frazer, Ballinger, Cox, Hinds, Stuve and Stewart2007).
We have enabled analysis of the X chromosome data, reflecting the increased analysis of this region in GWAS (Chu et al., Reference Chu, Shen, Xie, Miao, Shou, Liu and Huang2013; Conde et al., Reference Conde, Foo, Riby, Liu, Darabi, Hjalgrim and Skibola2013; Kou et al., Reference Kou, Takahashi, Johnson, Takahashi, Guo, Dai and Ikegawa2013; Tukiainen et al., Reference Tukiainen, Pirinen, Sarin, Ladenvall, Kettunen, Lehtimaki and Ripatti2014). Finally, we have made significant improvements in the analysis and data handling routines, increasing program efficiency.
Here we describe the VEGAS2 package, which is an extension of VEGAS with the ability to leverage the information provided by 1,000 Genomes phase 1 data, and allows gene-based analysis of the X chromosome.
Materials and Methods
Gene Data
We downloaded the hg19 annotated list of all RefSeq genes from UCSC table browser on May 22, 2014. After extracting genes located on the 22 autosomes and on the X chromosome, there were total 25,196 unique gene symbols; 5,356 symbols have variable transcription start and end positions. Symbols with overlapping transcription locations were merged to form a single full-length version of a gene. Cases where transcription sites were not contiguous with each other were given a new gene symbol with nomenclature ‘Originalgenesymbol_1/2/3’. In total, 26,056 unique VEGAS2 gene definitions (24,769 autosomal and 1,287 X-chromosomal genes) are used.
1,000 Genomes Data
VEGAS2 repository files were constructed using 1,000 Genomes phase 1 release version 3 was downloaded on May 22, 2014 from the NCBI website (ftp://ftp-trace.ncbi.nih.gov/1000genomes/ftp/release/20110521/). Using the vcftools package (Danecek et al., Reference Danecek, Auton, Abecasis, Albers, Banks, DePristo and Genomes Project Analysis2011), 1,000 Genomes phase 1 data were divided into the following ancestry groups: European (379 individuals), Asian (286), African (246), and Latin American (181). These four genotype datasets were filtered separately to extract SNPs with minor allele frequency above 1% and a Hardy–Weinberg p-value above 1 × 10−6. We also filtered out X chromosome SNPs that showed significant difference (p-value < 1.8 × 10−7) in allele frequency between males and females (there were 146 such SNPs in the European reference set, with similar numbers in other sets).
Gene-Based Association Testing Approach
In VEGAS2, the user has five options regarding gene boundaries for SNP selection:
-
1. SNPs within the gene, relative to the 5’ and 3’ UTR (0kbloc).
-
2. SNPs within 10 kb of the 5’ and 3’ UTR (10kbloc).
-
3. SNPs within 20 kb of the 5’ and 3’ UTR (20kbloc).
-
4. SNPs within 50 kb of the 5’ and 3’ UTR (50kbloc).
-
5. SNPs within gene plus any SNPs outside of the gene with r2>0.8 with SNPs within the gene (0kbldbin).
This allows the flexibility to include different sets of SNPs when testing for a gene-based association. Different gene boundary options have different advantages and limitations. For example, gene boundary option 1, ‘0kbloc’, focuses solely on intronic and exonic SNPs and ignores regulatory SNPs, reducing power if regulatory variation is important (and not tagged by SNPs residing in the gene). However, using a larger gene boundary may lessen the specificity of the result for a given gene because SNPs associated with neighboring genes may influence test statistics of a gene under consideration. SNPs a long distance from the gene are typically ignored in gene-based tests (Christoforou et al., Reference Christoforou, Dondrup, Mattingsdal, Mattheisen, Giddaluru, Nothen and Hellard2012) and so we have implemented gene boundary option 5, ‘0kbldbin’, to allow distant SNPs in high LD with genic SNPs to be included.
For each gene definition, the n SNPs’ p-values are first converted to upper tail χ2 statistics with one degree of freedom (df) and then summed to calculate a gene-based test statistic that would have a χ2 distribution with n df under the null hypothesis, if SNPs are in linkage equilibrium. Since linkage equilibrium for the n SNPs rarely occurs, their correlation is modeled using ∑, a n × n matrix of LD (r) values estimated from a 1,000 Genomes reference population. The user can choose a broad reference population group such as European (1000G EURO), Asian (1000G ASN), African (1000G AFR) and American (1000G AMR) using the option ‘-pop 1000GEURO/ASN/AFR/AMR’, or the user can choose a more specific population group with more similar LD to their population of interest. For example, the ‘-subpop GBR’ parameter can be used if the user wishes to calculate LD considering only individuals from the 1,000 Genomes reference population ‘British in England and Scotland (GBR)’. Significance is computed by comparing the summed χ2 statistics for each gene to simulated replicates from a multivariate normal distribution with mean = 0 and variance = ∑. Empirical p-values are computed for each gene using formula, p = r+1/m+1, where r is the number of instances where the simulated statistics exceed the observed data and m is the number of simulations.
We implemented a flexible gene-based approach where the user can specify what percentage of top SNPs are included in the gene-based test (the default is to consider all SNPs). This allows the user to include SNPs with more significant association with phenotype and remove SNPs that may dilute the summarized test statistics. An option is also provided to specify that only the single best SNP be included, which would be more relevant in genetic architectures where only few SNPs regulate the gene of interest and the top SNP is in high LD with those SNPs. A range of options is offered, since the best approach will vary depending on the true (unknown) genetic architectures.
We used the MD Anderson Cancer Centre melanoma cutaneous malignant melanoma case-control (MDACC-CMM-CC, 1,965 cases, 1,038 controls, typed on Illumina Omni-1M arrays) data (Amos et al., Reference Amos, Wang, Lee, Gershenwald, Chen, Fang and Wei2011) to compare VEGAS2 gene-based results obtained using only genotyped SNPs with results using 1,000 Genomes phase 1 imputed SNPs. We imputed chromosome 21 of the MDACC-CMM-CC data using IMPUTE2 software (Marchini et al., Reference Marchini, Howie, Myers, McVean and Donnelly2007) and performed association testing using SNPTEST (Wellcome Trust Case Control, 2007). VEGAS2 was applied to the summary results with and without imputation, using the default settings.
X Chromosome Gene-Based Test Approach
Although many commonly used genotyping platforms provide data on all chromosomes, relatively little attention has been paid towards analysis of the X chromosome in the GWAS setting. X chromosomes have some special characteristics compared to autosomes, namely:
-
1. males have a single copy; females have two copies.
-
2. one copy in females is fully or partly inactivated.
These special characteristics of the X chromosome require a separate statistical testing model for association analysis compared with autosomes. Different association testing models have been proposed (Clayton, Reference Clayton2008; Zheng et al., Reference Zheng, Joo, Zhang and Geller2007). Two popular models to analyze X chromosome GWAS data are: (1) sex-stratified (sexes analyzed separately) (Davidson et al., Reference Davidson, Cheong, Hysi, Venturini, Plagnol, Ruddle and Hardcastle2014; Zhang et al., Reference Zhang, Zhang, Yang, Wang, Zhang, Zuo and Yang2014); and (2) sex-combined, with X-inactivation modeled (males genotypes are coded as female homozygote, that is, males as 0, 2 and females as 0,1, 2) (Tukiainen et al., Reference Tukiainen, Pirinen, Sarin, Ladenvall, Kettunen, Lehtimaki and Ripatti2014). In a scenario where the proportion of males within cases is very different to the proportion of males within controls, the sex-stratified approach will have reduced power (Clayton, Reference Clayton2008). Hence, we suggest that users use the X-inactivation option as the default –– for example, input p-values from the default X-inactivation output from SNPTEST (Wellcome Trust Case Control, 2007). In addition to making the assumptions of X-inactivation and equal effect size in males and females (the per-SNP assumptions), VEGAS2 (by default) assumes LD and allele frequencies are equal across sexes. To minimize sampling error in this situation, LD and frequencies are estimated from both sexes combined. Users who do not wish to make these assumptions are catered for through the VEGAS2 ‘-sex’ option that treats each sex separately — in this case, the user should input separate p-values for the sexes separately. The sex-specific VEGAS2 outputs can be meta-analyzed using Fisher's method to combine the p-values.
We used MDACC-CMM-CC data (Amos et al., Reference Amos, Wang, Lee, Gershenwald, Chen, Fang and Wei2011) to test the X chromosome approaches in practice. First, per-SNP association was tested using SNPTEST (Marchini et al., Reference Marchini, Howie, Myers, McVean and Donnelly2007) using the X-inactivation model, with the p-values used as input to VEGAS2 (assuming similar LD and allele frequencies in males and females). Second, a logistic regression model in each sex separately was run, with the resultant p-values input into VEGAS2 with the ‘-sex’ flag specified, with the VEGAS2 output then meta-analyzed using Fisher's method.
Results and Discussion
Gene-Based Results on Different Sets of SNPs
We compared the gene-based results obtained using different sets of SNPs from MDACC-CMM-CC association data in chromosome 21 (Table 1). While using imputation fills in potentially informative untyped SNPs, on average the gene-based results do not differ dramatically when imputed and directly genotyped VEGAS results are compared (correlation 0.90 for total genotyped compared to total imputed). One advantage with imputation is that the number of genes with a gene-based result increased by ~25% for this data set (270 chromosome 21 genes covered with imputation compared with 215 genes with only genotyped SNPs).
Table 1 also shows the results with different levels of LD pruning of the imputed SNPs. Here, pruning means a SNP is removed if it has r 2 above the specified threshold with another SNP within a window of 50 SNPs as implemented in plink (Purcell et al., Reference Purcell, Neale, Todd-Brown, Thomas, Ferreira, Bender and Sham2007). A comparison between the gene-based result with r 2 > 0.99 and with no pruning is shown to investigate the phenomenon described by Moskvina et al. (Reference Moskvina, Schmidt, Vedernikov, Owen, Craddock, Holmans and O’Donovan2012). They showed that although the information content of the input data for ‘r 2 > 0.99’ and for ‘no pruning’ is similar (since the only difference is that one representative SNP is chosen each time two or more SNPs are in essentially complete linkage disequilibrium), the resultant correlation can be less than one. Table 1 shows that while we do see a correlation less than one, the high correlation (0.96) means that in practice the results will not differ substantially before and after pruning at this level. Examining pruning at lower r 2 thresholds, the unpruned and pruned results begin to diverge, as would be expected because the information content in the pruned set begins to decrease.
Since the information content of the input data for ‘r 2 > 0.99’ and for ‘no pruning’ is similar, there is unlikely to be an inherent advantage in considering the full set of imputed SNPs in practical applications of VEGAS2. Hence, in web-based version we implement ‘r 2 > 0.99’ pruning as the default in VEGAS2 (there is an option for the user to use no pruning if desired, although the runtime increases by four-fold). Specifically, when a user uploads their summary data, VEGAS2 first uses the user-specified 1000G reference set to remove all uploaded SNPs in r 2 > 0.99 with another uploaded SNP. The software then computes the gene-based p-values on the pruned set of SNPs. Similarly, in offline version user can provide pruned summary file as input to implement this method.
X Chromosome Gene-Based Test Using Sex-Stratified Versus X Inactivation GWAS Model
To test how the sex-stratified and X-inactivation models for GWAS on X chromosome behave in gene-based association test setting, we performed separate GWAS on MDACC-CMM-CC data using X-inactivation model on X chromosome using SNPTEST and run VEGAS2 using option ‘-sex BothMnF’ (default option). We performed association tests separately for each gender, and then ran VEGAS2 with option ‘-sex Males’ and ‘-sex Females’ respectively. We combined the gene-based p-values obtained from single gender analyses and compared it with the gene-based p-values obtained using X-inactivation analysis. As expected, the results from these two approaches are broadly similar, but given the different assumptions, not identical (Figure 1).
The gene PGRMC1 was more significant using the stratified sex model compared to the X-inactivation model (gene-based p-values, sex-stratified = 6.2 × 10−05, X-inactivation = 0.41). We further explored the results for the SNPs within this gene. This gene contains two genotyped SNPs, rs2499043 and rs11546862. Both these SNPs are significantly associated in the females-only tests, but not in males-only or X-inactivation tests (Table 1). Although the results in Figure 1 show reasonable concordance, the result for PGRMC1 illustrates that the assumptions made in the X chromosome analysis can in some cases greatly affect the results obtained using VEGAS2 (Table 2). In general, we recommend the sex-combined X-inactivation model, although users should be aware that in some cases the results may differ compared with the sex-specific model.
Web-Server Implementation
The online version of VEGAS2 is available through https://vegas2.qimrberghofer.edu.au/.
Offline Version for Linux System and Availability of Data Repository
VEGAS2 was developed in Perl programming language to work in Linux command line environment. The VEGAS2 data repository and scripts can be downloaded from https://vegas2.qimrberghofer.edu.au/zVEGAS2offline.tgz. The manual for installation and usage can be downloaded from https://vegas2.qimrberghofer.edu.au/VEGAS2usermanual.pdf.
Conclusion
In conclusion, we report on the VEGAS2 approach that uses 1,000 Genomes data to perform gene-based tests on GWAS summary results. VEGAS2 also extends the original VEGAS approach to perform gene-based testing on the X chromosome. Its offline implementation can be used in a Linux environment. The online implementation is publically accessible through the QIMR Berghofer webpage.
Acknowledgments
We thank Matthew Law for useful comments on the manuscript. We thank Xiaping Lin and Jonathan Davies from QIMR Berghofer IT for assistance with the VEGAS2 web application. AM is supported by an ANZ Trustees PhD scholarship. SM is supported by Australian National Health and Research Council and Australian Research Council fellowships. MD Anderson Cancer Centre (MDACC) melanoma case-control sample: The MDACC study is part of the Gene Environment Association Studies initiative (GENEVA, http://www.genevastudy.org) funded by the trans-NIH Genes, Environment, and Health Initiative (GEI). Genotyping of MD Anderson samples was performed through the University of Texas MD Anderson Cancer Center (UTMDACC) and the Center for Inherited Disease Research (CIDR), supported in part by NIH grants R01CA100264, P30CA016672, and R01CA133996, the UTMDACC NIH SPORE in Melanoma 2P50CA093459, as well as by the Marit Peterson Fund for Melanoma Research. CIDR is supported by contract HHSN268200782096C. The datasets used for the analyses described in this manuscript were obtained from dbGaP at http://www.ncbi.nlm.nih.gov/gap through dbGaP accession number phs000187.v1.p1. Principal investigators: Christopher Amos, PhD, University of Texas MD Anderson Cancer Center, Houston, TX, USA; Qingyi Wei, MD, University of Texas MD Anderson Cancer Center, Houston, TX, USA; Jeffrey E. Lee, MD, University of Texas MD Anderson Cancer Center, Houston, TX, USA.