An Evaluation of Statistical Approaches to Rare Variant Analysis in Genetic Association Studies By Andrew P Morris, Eleftheria Zeggini Published in: Genetic Epidemiology DOI: 10.1002/gepi.20450 ABSTRACT Genome-wide association (GWA) studies have proved to be extremely successful in identifying novel common polymorphisms contributing effects to the genetic component underlying complex traits. Nevertheless, one source of, as yet, undiscovered genetic determinants of complex traits are those mediated through the effects of rare variants. With the increasing availability of large-scale re-sequencing data for rare variant discovery, we have developed a novel statistical method for the detection of complex trait associations with these loci, based on searching for accumulations of minor alleles within the same functional unit. We have undertaken simulations to evaluate strategies for the identification of rare variant associations in population-based genetic studies when data are available from re-sequencing discovery efforts or from commercially available GWA chips. Our results demonstrate that methods based on accumulations of rare variants discovered through re-sequencing offer substantially greater power than conventional analysis of GWA data, and thus provide an exciting opportunity for future discovery of genetic determinants of complex traits. Genet. Epidemiol. 34: 188–193, 2010. © 2009 Wiley-Liss, Inc. INTRODUCTION Recent advances in whole genome genotyping technologies, the availability of large, well-defined population-based disease cohorts, and a better understanding of common human sequence variation, coupled with the development of appropriate quality control and analysis pipelines, have led to the identification of many novel common genetic determinants of complex traits [The Wellcome Trust Case Control Consortium, 2007; Zeggini et al., 2008; Barrett et al., 2008; Raychaudhuri et al., 2008; Cooper et al., 2008; Willer et al., 2009; Aulchenko et al., 2009; Prokopenko et al., 2009]. Nevertheless, despite these successes, much of the genetic component of these traits remains unaccounted for. Although there may be many undiscovered common polymorphisms associated with complex traits, it seems unlikely that the “common-disease common-variant” hypothesis is all encompassing. One unexplored paradigm which may contribute to this unexplained genetic component is a model of multiple rare causal variants, defined here to have a minor allele frequency (MAF) of less than 1%, each of modest effect, but residing within the same functional unit, for example, a gene. Joint analysis of rare variants within a gene, searching for accumulations of minor alleles within the same individual, may thus provide signals of association with complex phenotypes that could not have been identified through traditional association analysis of single nucleotide polymorphisms (SNPs), typically defined to have MAF of at least 1%. For example, minor alleles at multiple rare variants in ABCA1, APOA1 and LCAT have been demonstrated to contribute collectively to low plasma levels of high-density lipoprotein cholesterol [Cohen et al., 2004]. Currently, most studies of rare variants utilise data from commercially available GWA chips, which are far from ideal since they are designed for capturing common human genetic variation. However, the availability of data more appropriate for rare variant association analysis is just around the corner, with whole genome re-sequencing efforts, such as the 1,000 Genomes project (http://www.1000 genomes.org) soon reaching completion. Furthermore, large-scale deep re-sequencing technologies are becoming increasingly efficient and cost effective, and thus may soon be realistic for rare variant discovery in specific genes in large disease or population-based cohorts. We have developed a novel test of association with rare variants discovered through such re-sequencing efforts, based on the accumulation of minor alleles within the same functional unit, for example a gene-coding region extended up and downstream to incorporate additional functional elements and the regulatory region. We have then undertaken a simulation study to focus on two distinct, but timely, scenarios with the aim of addressing specific, as yet unanswered, methodological questions in each. First, when deep re-sequencing data are available to discover rare variants, do methods based on accumulations of minor alleles within the same functional unit offer greater power to detect association with complex traits than traditional analysis of SNPs on GWA chips? Second, when only GWA chip data are available, what is the most powerful strategy for identifying rare variant associations with complex traits? METHODS We consider two specific tests of quantitative trait association with accumulations of minor alleles across rare variants within the same functional unit. In the first of these tests, the phenotype is modelled in a linear regression framework as a function of the proportion of rare variants at which an individual carries a minor allele. In the second, the phenotype is modelled in the same regression framework, but this time as a function of the presence/absence of a minor allele at any rare variant within an individual. This collapsing approach has been previously proposed in the context of a binary trait [Li and Leal, 2008], and has been demonstrated to be powerful for detecting association with rare variants discovered through re-sequencing. Consider a sample of unrelated individuals, phenotyped for a normally distributed trait, and typed for rare variants in a gene or small genomic region. Let ni denote the number of rare variants for which the ith individual has been successfully genotyped, and let ri denote the number of these variants at which they carry at least one copy of the minor allele. We can model the phenotype, yi, of the ith individual in a linear regression framework, given by yi = E[yi]+ɛi, where ɛi,σE), and In this expression, xi denotes a vector of covariate measurements for the ith individual, with corresponding regression coefficients β. The parameter λ is the expected increase in the phenotype for an individual carrying a full complement of minor alleles at rare variants compared to an individual carrying none. As an alternative, we can model the expected phenotype of the ith individual as where I(ri) is an indicator variable taking the value 1 if ri>0, and 0 otherwise, in other words, the presence of at least one minor allele at any rare variant. Here, the parameter λ is the expected increase in the phenotype for an individual carrying at least one minor allele at any rare variant compared an individual carrying none. For either model, the likelihood contribution of the ith individual is given by We thus construct likelihood ratio tests of association of an accumulation of rare variants with disease by comparing the maximised likelihoods of two models via analysis of deviance: (i) the null model where λ = 0; and (ii) the alternative model for which λ is unconstrained. The contribution of the ith individual to the likelihood, f(yi | α,λ,β,ri,ni,xi), is weighted by ni to allow for differential call rates between samples. We denote the likelihood ratio test based on the proportion of rare variants at which an individual carries minor alleles by RVT1, and that based on the presence/absence of at least one minor allele at any rare variant by RVT2. Both RVT1 and RVT2 can be generalised to tests of association with a binary trait within a logistic regression-modelling framework. SIMULATION STUDY In order to evaluate the relative merits of different analytical approaches to identify rare variant associations with a quantitative trait, we have performed simulations using simple models of population genetics to generate high-density haplotype data in a 50-kb genomic region used to represent a functional unit of interest. We considered a range of models for association of the trait with multiple causal variants in the same region, under two different assumptions: (i) the mean trait value is determined by the presence or absence of a minor allele at any causal variant; and (ii) the mean trait value determined by the proportion of causal variants at which a minor allele is present. Trait association models were then parame-terised in terms of: (i) the maximum MAF of any individual causal variant; (ii) the total MAF of all causal variants; and (iii) their joint contribution to the phenotypic variance. Full details of the simulation process are described in the Appendix. We began by simulating a population of 40,000 haplotypes, and selected causal variants according to our chosen model of association. We selected 10,000 haplotypes, paired together at random to form 5,000 individuals in our “analysis cohort”, and generated their phenotypes according to their genotypes at the causal variants. We then selected a further 2,000 haplotypes in our “discovery panel”, used here to represent the deep re-sequencing data we expect from the 1,000 Genomes project. Over all simulations, the mean number of rare variants with at least two copies of the minor allele in the discovery panel was 52.2. We assumed that each of these rare variants was taken forward for genotyping in the analysis cohort, and tested for association using both RVT1 and RVT2. Our next step was to select variants in the 50 kb region to have similar properties to the Affymetrix Human SNP Array 6.0 in terms of mean density and MAF profile in the population of 40,000 haplotypes. All SNPs on the GWA chip were analysed independently, using conventional trend tests of association, with Bonferroni correction for multiple testing, and jointly, using standard haplotype-based techniques. Over all simulations, the mean number of GWA chip SNPs in the region was 14.8, while the mean number of SNP haplotypes with population frequency greater than 1% was 18.0. We attempted to apply RVT1 and RVT2 to the GWA chip data, but the mean number of rare variants in the region was just 0.2, and provided minimal power to detect accumulations of minor alleles at these loci. As a result, we extended our analysis to include “low-frequency” SNPs (1%