Quality control procedures for genome-wide association studies

Abstract

Genome-wide association studies (GWAS) are being conducted at an unprecedented rate in population-based cohorts and have increased our understanding of the pathophysiology of complex disease. Regardless of context, the practical utility of this information will ultimately depend upon the quality of the original data. Quality control (QC) procedures for GWAS are computationally intensive, operationally challenging, and constantly evolving. Here we enumerate some of the challenges in QC of GWAS data and describe the approaches that the electronic MEdical Records and Genomics (eMERGE) network is using for quality assurance in GWAS data, thereby minimizing potential bias and error in GWAS results. We discuss common issues associated with QC of GWAS data, including data file formats, software packages for data manipulation and analysis, sex chromosome anomalies, sample identity, sample relatedness, population substructure, batch effects, and marker quality. We propose best practices and discuss areas of ongoing and future research.

Figure 1

A flowchart overview of the entire GWAS QC process. Each topic is discussed in detail in the corresponding section in the text. Squares represent steps, ovals represent input or output data, and trapezoids represent filtering of data.

Figure 2

Visualization of X and Y probe intensities. The x-axis and y-axis represent the sum of the average over all probes for the normalized Cartesian intensity for allele A and the average over all probes for the normalized Cartesian intensity for allele B using all probes available on X chromosome and Y chromosome, respectively. The XX (female, red circles) and XY (male, blue triangles) subjects are shown on the bottom right corner and on the top left corner, respectively. The plot reveals two mislabeled individuals (one male with the female cluster, and one female with the male cluster). Several XXY individuals are also clearly visible (upper right corner).

Figure 3

Copy Number and allelic variation to detect anomalies on X chromosome. The top plot shows the B-Allele frequencies for all probes for one sample with total loss of heterozygosity (LOH) on X chromosome. The bottom plot shows the copy number variation from the same sample on X chromosome. Both plots are helpful to detect regions of LOH and/or copy number variation such as deletion and amplification.

Figure 4

Points in this plot show pairs of individuals plotted by their degree of relatedness: the proportion of loci where the pair shares one allele IBD (Z1) by the proportion of loci where the pair shares zero alleles IBD (Z0). These values are obtained from PLINK using the --genome option. Pairs are color-coded by the type of relationship determined by the pedigree information embedded in the pedfile (also reported by PLINK). This plot omits pairs of individuals having an overall kinship coefficient ≥ 0.05 for clarity. There is a pair of monozygotic twins represented by a point in the lower left at (0,0), because they share two alleles IBD at every locus across the genome.

Figure 5

Histogram showing the distribution of pairwise kinship coefficients (where kinship coefficient is greater than 0.05). The peak over 0.5 represents first degree relatives (parent-offspring, full siblings). The peak over 0.25 represents second degree relatives (half siblings, avuncular, grandparent-grandchild). Third and fourth degree relatives begin to blend into more distantly related samples between zero and 0.125.

Figure 6

Proportion of SNPs or samples remaining as call rate threshold increases. The green line shows the propotion of SNPs remaining when SNPs are discarded if they fall below the given genotyping efficiency threshold. The blue line shows the proportion of samples remaining, while the red line shows the proportion of samples remaining if a 99% call rate threshold is applied to eliminate poor quality markers first.

Figure 7

This shows the power to detect an association at genome-wide significance (p<5×10⁻⁸), assuming the actual causal SNP is genotyped in a case-control study consisting of 5000 cases and 5000 controls of a common disease with 10% prevalence under an additive model at several different odds ratios. Note that when the MAF is low, power is extremely low even for very large effects (OR=1.7).

Figure 8

AB and BB individuals are split into sub-clusters AB and AB’, BB and BB’, while AA cluster is unaffected. The AB/AB’ split results in some AB samples miscalled as AA (diagnosed by Mendelian inconsistencies in the genotypes), as well deviation from HWE due to excess homozygosity. Since only samples with at least one B allele demonstrate the splitting, one consistent explanation is the presence of a cryptic polymorphism near rs2301237 on a haplotype that contains the B allele. In this case, a second polymorphism (rs3114267) lies eight bases upstream from the typed polymorphism, and is in complete LD (D’=1, r^2=.2) with rs2301237.

Figure 9

Unexpected number of clusters resulting in departure from HWE consistent with copy loss. Hemizygous individuals cluster at AO and BO. Individuals with homozygous deletions cluster at OO and their genotype calls are missing. The AB cluster remains intact, since these individuals are ipso facto diploid at the locus. Parent-parent-child Mendelian errors are present when at least one parent is hemizygous and produces hemizygous offspring. The deletion results in excess homozygosity. In this case, the “copy loss” appears to be a six-nucleotide insertion (rs71578153) coincident with rs11591064 that disrupts both A and B probes.

Figure 10

The five observed clusters are most consistent with a segmental duplication, although none is curated around the locus. A copy number variant would be expected to produce additional clusters above the AA and BB clusters (ie, AAA and BBB), as opposed to the splits being confined to strictly the heterozygous clusters. Regardless, the artifact results in excess heterozygosity.