The effects of reference panel perturbations on the accuracy of genotype imputation
Reference-based genotype imputation is a standard technique that has become increasingly popular in large-scale studies involving genomic data. The two key elements involved in the process of genotype imputation are (1) the haplotype reference panel to which a target individual is being imputed, and (2) the imputation algorithm used to infer missing genotypes in the target individual. The imputation literature has historically focused mainly on (2), with a typical comparative study investigating the relative performance of various imputation algorithms while holding the reference panel constant. However, the role of the reference panel itself (1) on overall imputation performance is equally, if not more, important than the choice among many high-performing algorithms. Even though it is intuitive that the quality of a reference panel should play a role in the accuracy of imputation, it is nonetheless unclear to what extent common errors during panel creation (e.g., genotyping and phase error) lead to suboptimal imputation performance. In this study, we investigate the effects of applying three distinct modes of perturbations to a widely used haplotype reference panel in human genetics on the resulting imputation accuracy. Specifically, we perturb the reference panel by (1) randomly introducing phase errors, (2) randomly introducing genotype errors, and (3) randomly pruning variants from the panel (all at varying magnitudes). We then impute a set of diverse individuals at various sequencing coverages (0.5x, 1.0x, and 2.0x) to these various perturbed panels and evaluate imputation accuracy using the r2 metric for the entire cohort as well as ancestry-stratified subsets. We observe that both phase- and genotype-perturbations can dramatically affect imputation accuracy, particularly at very low allele frequencies, while pruning variants has a far smaller effect. We then empirically verified that our simulations reliably predict the impact of potential filtering techniques in a real-world dataset. In the context of haplotype reference panels, these results suggest that phasing and genotyping accuracy are far more important than the density of a reference panel used for imputation.