Efficient approaches for large scale GWAS studies with genotype uncertainty

1AbstractIntroductionAssociation studies using genetic data from SNP-chip based imputation or low depth sequencing data provide a cost efficient design for large scale studies. However, these approaches provide genetic data with uncertainty of the observed genotypes. Here we explore association methods that can be applied to data where the genotype is not directly observed. We investigate how using different priors when estimating genotype probabilities affects the association results in different scenarios such as studies with population structure and varying depth sequencing data. We also suggest a method (ANGSD-asso) that is computational feasible for analysing large scale low depth sequencing data sets, such as can be generated by the non-invasive prenatal testing (NIPT) with low-pass sequencing.MethodsANGSD-asso’s EM model works by modelling the unobserved genotype as a latent variable in a generalised linear model framework. The software is implemented in C/C++ and can be run multi-threaded enabling the analysis of big data sets. ANGSD-asso is based on genotype probabilities, they can be estimated in various ways, such as using the sample allele frequency as a prior, using the individual allele frequencies as a prior or using haplotype frequencies from haplotype imputation. Using simulations of sequencing data we explore how genotype probability based method compares to using genetic dosages in large association studies with genotype uncertainty.Results & DiscussionOur simulations show that in a structured population using the individual allele frequency prior has better power than the sample allele frequency. If there is a correlation between genotype uncertainty and phenotype, then the individual allele frequency prior also helps control the false positive rate. In the absence of population structure the sample allele frequency prior and the individual allele frequency prior perform similarly. In scenarios with sequencing depth and phenotype correlation ANGSD-asso’s EM model has better statistical power and less bias compared to using dosages. Lastly when adding additional covariates to the linear model ANGSD-asso’s EM model has more statistical power and provides less biased effect sizes than other methods that accommodate genotype uncertainly, while also being much faster. This makes it possible to properly account for genotype uncertainty in large scale association studies.