Scientist Stories: Marika Kaakinen, Genome wide association studies

Hello and welcome to the educational session 18 on introduction to the statistical analysis of genome-wide association studies and the first lecture, what, when, and whys of genome-wide association studies. My name is Marika Karginen and I'm a lecturer at the University of Surrey, UK. This educational session is based on the European Society of Human Genetics-funded short course that has been held since 2016. So far, 4 live courses and 2 online courses have taken place, educating over 270 scientists from around the world. On each course, there are a limited number of ESHE-funded fellowships available. The next online course will be held in January 2022, and more details can be obtained closer to the time from the University of Surrey and ESAG websites. This session has 3 talks. The first one aims to give an overview of GWAS, focusing on background and motivation for GWAS, then the steps and related to performing a GWAS and how to present the results from a GWAS study. Then some post-GWAS thoughts are given. The second talk considers issues related to genotyping, quality control, and imputation, and the third one focuses on population structure, meta-analysis, and transethnic meta-analysis. Prior to GWAS, twin studies, linkage studies, and candidate gene approach were the main statistical approaches in linking our genomes to human health and disease. Twin studies were mainly used to estimate heritability of different diseases and to dissect the role of genes from the environment. Linkage studies were conducted in related individuals and they are a hypothesis-free approach where the whole genome is scanned to look for genetic regions in linkage with the disease. In contrast to the hypothesis-free approach of linkage studies, candidate gene approach is hypothesis-driven, where the choice of the region that we are looking at is based on prior studies such as animal studies or understanding of the disease pathology, and then we call it the biological candidacy. Linkage studies in families had shown that they are highly useful for single gene disorders, but statistical power calculations had shown that they are not for common and complex diseases, which are often affected by multiple genes with small effects from each of them. Therefore, GWAS was proposed as a similar approach to linkage studies, but in unrelated individuals to increase the sample size. In that, we are also looking at the whole genome and looking if the allele of a genetic variant is found more often than expected in individuals with the phenotype of interest. And as said, they are better than linkage studies at detecting weak genetic effects. So why were GWAS studies not done already previously? They require a lot of things. So first of all, advances in genotyping technologies were needed to make so that they can analyze thousands of samples efficiently, and also this needed to be cost-efficient. Advances in software were also needed to make so that we can analyze hundreds of thousands or millions of SNPs for thousands, now hundreds of thousands of people. And also we needed computational capabilities to store all this data. And handle all this data that was coming from GWAS analysis. Biobanks were all of course also very useful in getting the data from the individuals to analyze. There are a couple of projects that need mentioning that have enabled GWAS. First of all, the International HapMap Project. Has identified a majority of the common SNPs, single nucleotide polymorphisms, that are analyzed in AGivos. The project has also identified haploblock structures so that we can focus on a subset of SNPs that describe most of the variation in the human genome. More recent projects such as the 1000 Genomes has continued the work from the HapMap project and the Haplotype Reference Consortium, and most recently the TOPMed project. A couple of key concepts that are essential in doing a GWAS study are, first of all, single nucleotide polymorphism, GWASes are focusing on these, the variation in a single nucleotide in human genome. Minor allele frequency is the allele frequency of the less common allele in the population, and these vary by populations, and GWASes usually focus on common variants with minor allele frequency over 5%. We are also looking at SNPs that are in Hardy-Weinberg equilibrium, so that the effects we see are not the effects from migration and so on. Linkage disequilibrium and haplotypes are also two important concepts that I will get into the following slides. So in GWAS, the idea is that we genotype SNPs across the genome, usually from, let's say, 500,000 to millions of SNPs, so that they represent a large proportion of variation in the human genome. Then we statistically test each of these SNPs against the trait of interest, It can be a continuous trait such as height, or it could also be disease phenotypes, for example, type 2 diabetes. Then, which SNPs to genotype? In January 2021, the dbSNP database listed over 900 million SNPs in humans, and this includes also rare SNPs. We don't want to genotype all of these, we want to be cost efficient, and using a so-called DAG SNP selection and the linkage disequilibrium that I mentioned, it is possible to cover all common variants using small subset of markers. For example, in Europeans it has been estimated that with 300,000 SNPs we can cover most of the common variation in the human genome. And in the figure here on the right, we see some haplotype blocks with high linkage disequilibrium, correlation between the SNPs, and by selecting just 1 or 2 SNPs from each of these haplotype blocks, we can cover the whole block. There are mainly two commercial arrays, Affymetrix and Illumina, on the market, and their amount of SNPs ranges usually from around 300,000 SNPs up to 5 million SNPs. Then there are also arrays that focus on specific SNPs, such as those related to metabolic diseases or immune function, or arrays that genotype only the exomes of the human genome. InChivo studies are usually analyzed in multiple samples and then meta-analyzed to increase power for the study. But now you may wonder that if different studies have used a different array, they have different SNPs in their arrays. So to enable meta-analysis on the same set of variants, imputation of the untyped SNPs is done, and more on this in Lecture 2 on this educational session. One issue with genome-wide association studies is multiple testing, because we are running hundreds of thousands or even millions of regression analyses. So we need strong evidence for each SNP to avoid false positives. In our findings. And one of the most commonly used corrections is the Bonferroni correction, by dividing the original level of significance by the number of tests we are performing. And if we assume 1 million tests, which is coming from 1 million independent SNPs in the human genome, we get a new level for statistical significance of 5%, 10^-8. There are also other approaches such as the false discovery rate and bias factor, but Bonferroni correction is the most commonly used, although somewhat criticized for conservativeness. The first successful GWAS was reported in 2005 for AIDS-related macular degeneration. It had a fairly small sample size of only 96 cases and 50 controls, but because of the strong association, it identified 2 SNPs on chromosome 1, which can be seen on this first-of-its-kind Manhattan plot, where we have on the x-axis, we have the SNPs and their positions, and on the y-axis, we have the minus log 10 p-value. The lines here represent the association signals from each of these SNPs, and we can see the strong association on chromosome 1. The first large-scale genome-wide association study was reported in 2007 by the Wellcome Trust Case-Control Consortium. They analyzed 7 diseases and used a pooled set of controls, and they successfully reported associations for most of the diseases across the genome. They also introduced new software and analysis methods for analyzing GWAS, and many studies still today follow this protocol. And also use the software introduced in this study. Since their introduction in 2005, genome-wide association studies have reported a number of association signals for a number of diseases and across the genome. From this figure, we can see the different chromosomes, and in 2005, we see the one association on chromosome 1. But when we go to 2020, we see that most, for most of the chromosomes, we have reported multiple associations for multiple different traits, and these are colored here with the different color codes. When starting to conduct a genome-wide association study, it is very important to evaluate the statistical power of your study. The power is the probability that the test will reject a false null hypothesis. As with any epidemiological study, factors that affect power are the size of the effect, so we are more powered to see larger effects, and also the standard deviation of the characteristic we are looking at. The more spread it is, the less powered we are. Bigger sample size of course increases the power, and the smaller the significance level is, the lower the power. And one more additional thing we need to consider in GWAS is the allele frequency. We are more powered to find effects for common variants than for the rarer variants. Quality control is an essential part of genome-wide association studies. It is often the most time-consuming part, but it is also the most important so that we can avoid systematic bias in the results. It is done on two different levels: on sample and SNP level. And in samples, we are checking that there are no sex mismatches, outlying missing genotype or heterozygosity rates. We are looking at duplicates and unknown related individuals in our data, ancestry outliers, and also the genotyping quality. For each individual. For SNPs, we are also looking at the genotyping quality and deviation from the Hardy-Weinberg equilibrium. The statistical analysis in genome-wide association studies starts with defining the genetic model, and in GWAS, the most commonly used model is additive. So this means that we are looking at the genotypes and their risk alleles, and we are coding for each individual the number of copies of the risk allele they have. And then we use this number of copies, either 0, 1, or 2, as a continuous variable in the analysis. And as we can see in this example figure here, For body mass index and the FTO genotype, we see that individuals with more copies of the AA risk allele have an increased body mass index. And of course, there are other types of genetic inheritance, but this additive has been shown to capture also them very well. The statistical model itself depends on the trait that you are studying. If you have a quantitative trait, you can do a linear regression, but if you are analyzing a disease phenotype, then you would obviously do a logistic regression, or you could do survival analysis and so on. As in any statistical modeling, we can adjust for covariates, Usually we have some study-specific covariates including age, sex, body mass index, and so on. And one important covariate is the population stratification, and more about this on the third lecture. And in this example here, we see a simple model for body mass index where we are associating each SNP at a time on each and have two covariates, age and sex. Genome-wide association studies are usually done in collaboration and often internationally, and this is to increase the sample size and the power for the study. And before the analysis, a common study protocol is agreed between the participants to harmonize the analysis as much as possible to avoid bias. And this harmonization usually covers phenotype coding, any exclusion criteria, and any covariate adjustments. The analysis stage usually consists of 3 different stages: the discovery stage, where several independent samples are analyzed and then meta-analyzed, and then the results from this discovery stage are replicated in independent samples so that we can separate false positive associations from true positive associations, and the replication sample should be as similar as the original discovery sample, so for example from the same ethnicity. A lack of replication may be due to unaccounted population stratification, some phenotype differences, selection biases, or genotyping errors. So genotyping quality control is very important. And in the end, usually the discovery and replication samples are meta-analyzed. The first plot you usually do when you perform a GWAS is the QQ plot, and this is a diagnostic tool where you, where we plot the observed p-values or test statistics against expected, and we expect the observed line to follow closely the expected line.. But if there is deviation, even with small values, then it indicates that there's some unaccounted population stratification that should be dealt with before proceeding with the analysis. Deviation at the larger values indicates some actual associations, as we can see in this example plot. So we see some low p-values, as presented by the minus log p-value here. The second plot that we usually do is the Manhattan plot, and on the x-axis we have the chromosomes and positions, and on the y-axis we have the minus log 10 p-values, and each dot represents here an association result for each SNP. SNP. So in this plot, we see association results from a birth weight GWAS, and we have highlighted here some potential associations in red. These are the associations that reached the threshold for genome-wide significance of 5 × 10����. And after we have identified these signals, we usually take a closer look at each of these signals in more detail. Here we are looking at the CDcol1 region from the previous slide in more detail by using a so-called regional plot or locus zoom plot. And here we see the most significantly associated SNP presented by its rsID, and then we see the linkage disequilibrium, the correlation with other SNPs in this region, presented by the color scheme shown on the left-hand side of the plot. And we see that there is some high correlation between this SNP and other SNPs in in the region, so we cannot be sure which one is actually the causal variant in this region. So some follow-up analyses are needed to find the causal variant from this region. When the GWAS is done in collaboration, we usually also show a forest plot to show the results by cohort to see if they are in alignment or if there is heterogeneity between the cohorts. So here we see for this same birth weight analysis and this same CDKL1 region and one specific SNP from that region, we see the effects by different cohorts and then We see the overall effect from discovery sample and the overall effect from the replication sample, and then the final overall combined effect, which shows a negative effect from this SNP on birth weight, indicating that having this SNP lowers your birth weight. Finally, results are of course presented also numerically in tables, and they quite often show the different stages: the Stage 1, the discovery stage, Stage 2, the replication stage, and then the combined stage results. And for each association signal, the locus name, usually the nearest gene is shown, and then its chromosome and position, the rsID, the effect allele and its frequency in that population, then the effect estimate and the associated p-value. Once a significant signal has been identified there are 3 possibilities for that SNP. Either this is the causal variant for that trait of interest, or as we saw from the locus zoom or regional plot, it could be in linkage disequilibrium with the actual causal variant. Or the final possibility is that it is a false positive finding, when replication studies become very important. And genotyping quality control issues are also very important in avoiding systematic errors. But once we have identified our signals, we need to provide biological evidence for the identified signal and try to find the causal variant. How to do these things is: we could do fine mapping of the region by sequencing or statistically. So for example, by statistical fine mapping, we could try to identify the so-called credible set of causal variants in that region, or we could use some statistical software such as FineMap and Caviar. We can do biological annotation by using multiple databases that are available, and we can also do functional annotation by looking at gene expression and epigenetic annotation. We could look at biological pathways by using some softwares and databases. Then we could go on and try to calculate the explained variance for our phenotypes and polygenic risk scores, Further on, we can do causal analysis by using Mendelian randomization or some other follow-up analysis such as stratified analysis and interactions. As we saw from the figures, the GWAS has been incredibly successful for many of the traits, but still they explain only a small proportion of the estimated heritability for many of the traits, and this missing part is called the missing heritability. There has been some discussion that maybe the heritability estimates in the first place are biased, and there are new techniques in calculating the heritability estimates. But anyway, we need to try to find the missing heritability, and we could try to do this to increase sample size, Biobanks have been very useful in this. We could do gene and environment interaction analysis and gene-gene interaction analysis. We could look into the rare variants or use novel analysis strategies to increase power. To finalize this talk, I have listed here some selected software and some databases and also some selected references to get you started with on your journey in GEOS.