GWAS is a key workflow of the H3A. Our recommended approach to SOP is shown below. Note, this is not an algorithm – there are multiple ways of doing QC which are good. Also, what is shown below is not necessarily linear. For example, some of the analyses at the plate level might only be possible after doing an initial QC and then coming back to look at plate and batch effects.
While this SOP is designed for bioinformaticists doing the GWAS analysis once the data are ready, we recommend that it be studied thoroughly to help plan the entire process. This GWAS is written as a general guide for bioinformaticists, and in particular to assist groups undertaking H3ABioNet accreditation, both to prepare for and do the accreditation exercise.
Disclaimer: * Although we hope this SOP is educational and will help groups learn to do GWAS, it is not meant as a tutorial or a complete checklist. *Groups undertaking accreditation should realise that accreditation is undertaking by an independent international evaluation committee, and while they will have regard to the SOP, they make their decision at their own discretion.
Tool and file format
This SOP assumes that the data are in the binary format used by the PLINK software suite. However, this is not mean to say that PLINK is the only or even best tool to be used. The PLINK binary format (hereafter referred to as bped) encodes a dataset as a set of three files, with the following suffixes to their names:
- .bed: this is the binary genotype data, stored as 2 bits per genotype per sample
- .bim: a text file containing marker information, one line per marker, in genomic order
- .fam: a text file containing sample pedigree information
Glossary of associated terms and jargon
- Case/control: A type of GWAS in which all individuals in the study are placed into one of two groups. Case has the disease or trait of interest, and control does not. Alternatively, GWAS can be performed on a quantitative trait.
- dbSNP: A database maintained by NCBI at https://www.ncbi.nlm.nih.gov/snp/ where metadata about variants can be found.
- Genotype calling: The process of determining the genotype of each individual in the population at each variant of interest.
- GWAS: Genome-wide association study. An analysis that identifies genomic variants that are associated with a disease or trait in a population, and are putatively linked to a causative variant.
- Illumina: A company that produces genomic arrays and DNA sequencing equipment and reagents. In this SOP, we are generally referring to Illumina genotyping arrays.
- Imputation: Filling in missing data using estimated values. In GWAS, SNP data are commonly imputed using known linkage disequilibrium with nearby SNPs.
- Minor allele frequency (MAF): A statistic for each SNP, ranging from 0 to 0.5, indicating the frequency of the less common allele in the population. This is out of the total number of allele copies in the population, so if there were two heterozygotes and no minor allele homozygotes in a population of 100, the minor allele frequency would be 0.01.
- PCA: Principal components analysis. In GWAS, it is generally used to reduce thousands of SNPs to a few variables that describe population structure.
- PLINK: A popular software for performing GWAS.
- Population structure: Generally resulting from geographic separation of subpopulations, population structure means that different subgroups of the study population have different allele frequencies. Population structure can be a confounding factor in GWAS if the disease or trait is more common in some subpopulations than in others.
- SNP: Single nucleotide polymorphism. A common type of variant used in GWAS. SNPs are straightforward to assay from sequencing or array data.
The raw output of the genotyping process are files containing two signals for each genotyped sample – for Illumina products, these are IDAT files. The two signals correspond to the two alleles for a single SNP. These are commonly visualized as a scatter plot with one signal on the x-axis and the other signal on the y-axis, with each individual will be having one dot on the image. For diploid SNPs, typically the dots will visually cluster into three groups: those homozygous for the reference allele, those heterozygous, those homozygous for the alternate allele. Note that the IDAT files just describe the raw assay data of each individual for each SNP: usually the clusters are visually obvious, and the process of calling is to rigorously put each call in one of three clusters.
Typically most SNPs will have well behaved clusters and most individuals will clearly associate with one of the clusters, but there are always errors, anomalies and some borderline cases. Different clustering algorithms will produce different results. Typically, genotyping centres will provide a cluster report which describes for each SNP and each individual what the calls are. While well reputed centres will produce high quality calls, it is always a good idea to try alternate approaches. Well known calling tools are Illumina’s GenomeStudio and crlmm.
Our SOP does not yet take this process forward.
As is noted later, part of the QC will require coming back to the raw data.
Sources of error
Error can occur at multiple places, and these errors can easily introduce false signals or mask real signals.
- poor record keeping in the field, in transport, in storage, DNA extraction, plating and shipping
- sample contamination
- DNA extraction
- Problems with plates
Effort should be made to detect and eliminate any such problems.
Problems may occur randomly but there may also be due to non-random effects. For examples, a problem with a particular consignment of reagents may make two or three days’ worth of DNA extraction unreliable. The better record keeping there is of samples, the easier it is to work out problems. Particular considerations are listed below.
Plate Level errors
Currently (2019), samples are processed in batches of 96 wells (8x12), with one sample per well. Comparative QC should be done across the plates according to the QC parameters described below. If an entire plate is found to behave poorly, it should be removed.
A batch is a set of samples that are processed together. Ideally for consistency, all samples would be processed at the same time (collected, DNA extraction, genotyping) but for any real study this is not practical. As an example, in one H3A project, the samples were divided into four roughly equal batches, each batch corresponding to one campaign of DNA extractions. Batches 1-3 were shipped together to the genotyping centre and genotyped, and about 6 months later Batch 4 was shipped and genotyped. In this case, the group had to compare the four batches with each other and compare batch 1-3 with batch 4.
There may be different ways of dividing samples into batches (e.g. by recruitment stratum, when DNA extracted etc).
If participants are recruited in different places, differences may occur due to the way in which samples are handled at these places.
Case/control or phenotype differences
Of course, the whole purpose of a GWAS is to discover genetic differences between cases and controls. However, in a well designed study it very highly unlikely that the condition or disease of interest will be associated with widespread genomic differences. If there are overall genomic differences then this is most probably due to the way in which participants were sampled or the different processes in which DNA was extracted or genotyped. For example, a study might recruit 3000 cases and use results from a matched cohort for controls. As different people/labs might be responsible for collection and processing of the cases and controls, even if identical arrays are used and genotyping is done at the same centre, there may be systemic errors.
One useful phenotype to use for QC is sex. For many studies, it would not be expected that in the autosomal SNPs that there are differences at the genomic level.
As discussed later, population structure is a potential confounding factor. This may also make QC more difficult. In a highly homogeneous sample, one would not expect there to be significant genomic level difference between plates (e.g., when doing a PC analysis). However, in a large multi-country project with diverse ancestries, this would not be unexpected. This is especially because unless a deliberate effort is made to randomise (which may be not be practical) it is likely that batches of samples from the same site are likely to be processed on the same plate.
A particular concern in a study with population structure is if the number of cases and controls is not balanced across different sub-groups, and even more so when differences in both environmental and genomic background might cause different variants to be associated with the trait in different sub-groups.
Many errors will be due to random effects at either the SNP or individual level. The systematic errors above are often detected by comparing groups of things with each other, while random effects are usually detected using some experiment-wide cut-off, for example using the GenCall value (see below).
Evaluating sources of error
To test the various sources of error listed in the previous section, one approach would be to perform a principal components analysis (PCA) of the SNP data, and plot either the first two axes, or some number of axes determined by examining a scree plot. Every individual would be represented by one point in the PCA, and the points can then be colored by factors such as plate, batch, site, and case/control. Ideally in a randomized experiment, the observed clustering will not correspond to any of those factors. The PCA should also be plotted with individuals colored according to their reported ancestry, which should largely correspond to the observed clustering. If clusters are observed that don’t seem to have anything to do with ancestry, care should be taken to determine if they correspond to any known experimental factors.
Another useful approach would be to calculate the median or 10th percentile GenCall score per individual (see below) and relate that to various experimental factors. For example, a box plot or violin plot could show the distribution of median GenCall scores for each plate or batch. One should consider removing any poor quality plates or batches from the association analysis.
For Illumina array data, the key quality measure is called GenCall. Once clustering is done (recall, clustering is done per SNP, using the (x,y) coordinates of all the individuals for that SNP), a GenCall score is given for each sample/person of that SNP which a confidence score of the correctness of the call. This confidence score is a function of the clustering algorithm used. Illumina indicates that GenCall scores above 0.7 indicate highly accurate genotypes, and scores below 0.2 indicate failed genotypes. From this is derived
- The call rate or call frequency. The person doing the calling picks a minimum acceptable GenCall value, and any calls lower than this are rejected. The call rate is the proportion of calls that were not rejected for a given SNP or individual.
- The 10%_GenCall value: The GenCall value of the SNP at the 10% percentile of GenCall value (e.g, the SNP/individual for which 90% of other individuals have a better GenCall score). For a well genotyped SNP or individual this should be a very high quality absolute value (e.g. 0.7) even though it is relatively poor.
These can be measured per sample or per SNP. Good graphs to look at are call rate versus sample number, and call rate versus 10%_GenCall value, both of which may be used to identify poor performing individuals.
If biological replicates have been included in your study (a good idea), the concordance between these replicates should be examined as these will give good empirical evidence of the quality of genotyping. Similarly, if there are big differences in genotyping for biological replicates, these may indicate problems in sample handling.
The following should be checked in the QC process:
- Missingness at the SNP-level. SNPs with high missingness (i.e. a low call rate) should be removed. Usually this is set at the 1-2% level. (PLINK
- Missingness at the individual level. Individuals with high missingness should be removed. Usually this is set at the 1-2% level. (PLINK
Of course, the two are inter-related so a small benefit may be gained by iteratively removing badly behaved SNPs and poor individuals. Note that if a significant number of poorly genotyped samples are found, it may be desirable to remove the individuals and re-call the genotypes since poorly performing samples may significantly affect how clustering is done.
- Missingness at the SNP-level. SNPs with high missingness (i.e. a low call rate) should be removed. Usually this is set at the 1-2% level. (PLINK
Sex concordance. Sex concordance is typically used as a proxy that there hasn’t been swapping of samples. We check that the individual’s biological sex as recorded in our meta-data about the individual matches what the individual’s genotype says, typically as measured by the computing the inbreeding co-efficient of the X-chromosome. PLINK computes this with the F-statistic. By default, individuals with an F statistic less than 0.2 are regarded as female, and individuals with F-statistic greater than 0.8 are male. Of course the scores are 0.2 and 0.8 are arbitrary so care must be taken. In large samples, there may be people with unusual F-statistics for a variety of reasons. However, if a project has an individual recorded as male and their F-statistic is 0, then the most likely explanation is that there has been sample mishandling at some stage and that results we have are not for the individual we think it is. Any individuals who fail the F-statistic like this should be removed. This test is a proxy for testing correct labellings and record keeping: it cannot detect swaps where M/M or F/F swaps have taken place. If p sex concordance errors are found, it is likely that there are p other swap errors that have not been detected. Pay attention to the distribution of errors – are they randomly distributed or bunched in batches or groups. In some extreme cases it may be necessary to remove an entire group.
Population structure concordance. Using population structure approaches (discussed below), individuals should cluster appropriately. Some insight and judicious judgement is required depending on the sample chosen and what is known about the individuals. The more meta-data, the more accurate our testing can be done. But for example in a multi-centre trial, if an individual has told us that all four of her grand-parents were Zulu-speaking and identified as Zulu, but genetically the individual clusters with other individuals from a West African site then (although there are all sorts of interesting life stories) there is a strong possibility that this is a sampling error. As discussed above, population structure differences between batches or plates may also indicate errors, but in a heterogeneous study or multi-centre study this may be hard to test for. Individuals whose PC position falls far from the centroid of the group are likely a result of mislabelling or even genotyping problems.
Minor allele frequency (MAF). The ability to reliably detect and quantify variation is limited by the sample size and error rate of genotyping. A minimum minor allele frequency must be imposed since with very low MAF even a low error rate in genotyping can cause cases to be appear different to controls. Additionally, SNPs with low minor allele frequency have low power to detect associations, but still impact multiple testing correction, reducing confidence in associations detected at other SNPs. SNPs should be removed that do not meet or exceed the minimum minor allele frequency.
Difference in MAFs between sub-groups of the study (either tested directly using chi^2 or similar test or PC analysis) may also indicate batch effects of different sorts. As mentioned above this is complicated by population structure.
- Deviation from expected heterozygosity. The theory of Hardy-Weinberg equilibrium indicates that if the MAF of a SNP is f, the distribution of [homozygous minor allele, heterozygous, homozygous major alleles] should be [f^2, 2*(1-f)f, (1-f)^2]. Deviation from this may imply problems with the array. Batch effects of different sorts should be considered. Deviations from heterozygosity can be detected at the SNP and sample level. However, there are complexities.
- in a recently admixed population, deviations from HWE are likely;
- SNPs implicated in a disease may well be out of deviation from HWE, especially in cases;
- Typically we test for deviation from HWE by performing a chi^2 or similar test for each SNP. Since we may be testing hundreds of thousands or millions of SNPs we may need to adjust for multiple testing.
Thus there are reasons where deviation from HWE may not be a sign of error; but deviation from HWE may also indicate a problem with the assay. This makes setting the correct cut-off value for HWE testing a challenge.
--hweoptions will be very useful.
- Relatedness. There are several ways of measuring relatedness and you should consider the most appropriate for your study. One way of measuring relatedness is the PI-hat score computed by PLINK. There are two reasons for testing for relatedness (the discussion below uses PI-hat, but similar considerations apply to other approaches, mutatis mutandis).
- If there are pairs of individuals with a PI-hat score close to 1 then this indicates (a) a known biological replicate; or (b) a sample handing error; (c) the same individual being recruited more than once into the study. In case (a) one of the pair should be removed; in case (b) and (c) it is likely that both pairs will need to be removed. PI-hat values significantly above 0.5 but less than 1 are highly unlikely biologically and if you find any such pairs these are probably caused by contamination of some sort.
- The level of relatedness in the study, which will determine analytic approach. For family based studies a high level of relatedness is expected and it is important to check that the genotype-computed relatedness matches what is expected in the study. For most GWASes samples are ideally randomly recruited and some methods rely on low levels of relatedness. For samples recruited in large urban areas, random sampling should result in low levels of relatedness and if there are a significant number of related pairs, there may be a problem in experimental design and execution. In smaller regions, depending on region history and cultural practices, a relatively high number of related pairs is possible.
If methods such as PLINK’s chi^2 or linear or logistic regression are used, a cut-off value of relatedness should be used and one member of each pair with a PI-hat value higher than this cut-off should be removed. It is hard to say what a correct cut-off value is – the literature has studies where PI-hat values ranging from very low (e.g. 0.01) to 0.2 and the choice may be an expedient one. Recall that PI-hat of 0.125 corresponds to first cousin relatedness and 0.25 to grandparent-grandchild relatedness. There are several studies that use 0.18 as a cut-off. Higher than this may be problematic.
Alternatively, a GWAS method that can handle relatedness should be used. For example, mixed models approaches such as GEMMA and BOLTLMM can handle related samples if used appropriately.
Pairwise relatedness can be computed using PLINK’s
- Samples with chromosomal abnormalities Chromosomal abnormalities such as aneuploidy or long stretches of homozygosity should be tested for and the causes identified. These cases should be examined with a geneticist to determined possible causes and decide whether the sample should be removed.
As a byproduct of the genotyping technology, and the annotation data that accompanies the chips, SNP data from both Illumina and Affymetrix platforms may be reported as the allele on either the “forward” or “reverse” strand. Although the information to orient these calls properly is contained within the annotation data, conversion from the native format to PLINK format can be tricky, and errors will not be evident in “ambiguous” AT/GC SNPs. Therefore it is prudent to check that the alleles reported in the bim file match the known alleles in dbSNP or on the Ensembl genome browser. Another easy check is to identify the control samples that are often included in the dataset (eg samples from HapMap) and compare these to their known genotypes.
This is a particular problem when you will be merging your data with other sets.
Merging data from different genotype experiments
The following document describes some recommended steps for processing genotype data from a genome-wide association study, as will be generated by many of the H3Africa projects. Although it is written with data from SNP genotyping arrays in mind, many of the steps also apply to full genome sequence and exome sequencing data. Many of the tests can be performed with the PLINK software suite, but where other software is required this will be indicated.
Some steps below may be repeat QC steps mentioned above. If QC has been done thoroughly above then they may be omitted now.
Population stratification refers to structure within the sample group that is the result of systematic genetic differences between individuals that correlate with the phenotypic data. This could result in allele frequency differences that are due to ancestral proportion differences between cases and controls being mistaken for an association with the phenotype.
Population structure be can used
- to detect errors (as describe above)
- remove individual outliers – these are not errors but individuals whose genetic background are so different from others in the group that they may confound result
- adjust for genetic differences between and within groups
Allele frequency based methods, such as STRUCTURE or Admixture, can be used. These methods determine ancestral populations (which may be real, or may simply be a useful model for describing the data) and the allele frequencies of those ancestral populations, then assign each individual a proportion of ancestry to each population. The ancestry proportion matrix is commonly referred to as a Q matrix. Typically the method should be run for a range of different values of K (number of ancestral populations), and likelihood scores are used to determine the optimum K. To help the method identify each ancestral population (particularly ones that might not be well-represented in the dataset), it is helpful to add SNP data from known reference populations such as 1000 Genomes population.
PCA based methods require less computational time than allele frequency based methods, although they might not capture complex patterns of ancestry quite as well. One strategy is to use Eigensoft’s smartpca, and examine the top n eigenvectors. If any of these eigenvectors display segmentation based on a potential confounder, these factors should be examined further, and possibly corrected for. Individuals that differ significantly from the rest in terms of cluster membership should be investigated for possible removal. A common threshold in Eigensoft is to remove individuals that are outliers by more than 6 standard deviations.
Both these methods should be used on the data in an exploratory fashion, to try identify possible confounding factors and to understand how to handle population structure in the GWAS.
Different types and levels of structure can be observed. At one extreme, sampling from a very homogeneous group, relatively isolated from others will lead to a very tight cluster observable in a PCA. In such a case, population structure may be ignored. In an admixed group formed 10 generations or so ago in which random mating has taken place, a single cluster may be observed that is fuzzy and loose. In a GWAS recruited from multiple sites, multiple distinct clusters may be observed with few intermediate individuals, but more complex patterns can be seen, particularly in cosmopolitan, African regions. Population structure can be managed by adjusting
- GC control
- PCA analysis: including the right number of eigenvectors in the analysis
- Using mixed models approach
- Some combination of the above
Autosomal, X, Y and MT SNPs
Many GWASes only consider the diploid chromosomes and so only autosomal SNP are considered.
Association testing for single locus
The most common tests are single locus testing. Most GWAS tools provide multiple tests and insight into your biological question is needed. For example, if it is known that the disease being studied is recessive or dominant the correct variation on the association test can be used. A disadvantage of single-locus approaches is that they tend to overestimate the proportion of phenotypic variance explained by each SNP.
Some of the approaches are
- chi^2, logistic regression (case control)
- linear regression (quantitative study)
- mixed models approaches (e.g., GEMMA, BOLTLMM) which
Multiple testing should be dealt with, for example using Bonferroni correction, Benjamini and Hochberg’s false discovery rate (FDR), or permutation testing.
Covariates include population structure, sex, smoking, age, socio-economic status, etc. How to handle covariates is complex. Crudely there are two types of covariates:
- confounding covariates where not including the co-variate may cause false signals. Population structure is a good example of this. These are often covariates which mediate the genomic background (like population structure)
- covariates where not taking them into account will hide signals
Where possible (depending on the software used), covariates can be included directly in the model for association testing. An alternative approach would be to regress the trait of interest on the covariates, calculate the residuals, and then use those residuals as a quantitative trait in GWAS.
Imputation can be done in two modes – first, one can use imputation just to fill in missing genotypes. Second, imputation can be done to impute values for as many positions as possible, including SNPs that were not genotyped on the array but are part of known haplotype blocks. The latter is desirable when combining data across multiple genotyping platforms.
There are several approaches that iteratively add SNPs to the model until a sufficient amount of phenotypic variance is explained. One advantage of this is that when multiple SNPs are all linked to the same causative locus, only the top (i.e. most tightly linked) SNP will be added to the model, and the rest will be omitted because they do not explain any additional phenotypic variance. Each SNP that is included in the final model represents a distinct putative causative locus, and the proportion of variance explained by each SNP will not be overestimated as it is in single locus approaches.
The relatively low power of most GWAS designs means that a substantial number of false positives can be expected to be generated, so validation via replication in an independent population is an important part of most studies. A common strategy is to genotype a subset of the samples on a high coverage array, and then follow up with targeted genotyping of markers that appear interesting, on a larger number of samples.
Related to the issue of replication, meta analyses can help to distinguish true associations from false positives. This approach involves searching across multiple published studies to identify genomic loci that have repeatedly been found to be associated with a given trait. Linkage disequilibrium is an important consideration here; different SNPs within the same haplotype block can be considered as evidence for the same causative locus.
Practice dataset for GWAS studies can be found here
Highland et al, 2018. Quality Control Analysis of Add Health GWAS Data. https://cdr.lib.unc.edu/concern/parent/1n79h642n/file_sets/g158bk38b (Example of a good QC)
Illumina Inc, 2014. Infinium Genotyping Data Analysis. https://www.illumina.com/Documents/products/technotes//technote_infinium_genotyping_data_analysis.pdf
Turner et al. Quality Control Procedures for Genome Wide Association Studies. Current Protocols in Human Genetics, 2011,
Zeggini and Morris 2010. Analysis of Complex Disease Association Studies: A Practical Guide. Academic Press.
This work is licensed under a Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License.