Benchmarking imputation accuracy in the presence or absence of a reference panel

Whole genome sequencing (WGS) of a large number of samples is costly. Solutions to reduce this cost include targeting a proportion of the genome (e.g. SNP arrays) or lowering the sequencing depth (low-coverage WGS, lcWGS) but both solutions suffer from either genotype missingness or uncertainty. Genomic imputation addresses this problem by inferring missing or uncertain genotypes using a collection of high quality genomic data (a reference panel). However, certain methods can impute a lcWGS dataset without a reference panel. Because the investment into generating a reference panel can be prohibitively expensive in non-model species, a benchmarking of the accuracy of these alternative methods of imputation can help inform study design. Here, we imputed a dataset of 2800 lcWGS in the presence and absence of a reference panel of 502 samples. We used 32 individuals sequenced at both high and low coverage to estimate the accuracy of each method and explored the limitations of lcWGS sample size and reference panel size. Although the best results were achieved with a large reference panel, using only a lcWGS dataset showed accurate imputation, when over 500 samples were used and we account for missing data and low frequency alleles. In addition, imputing with or without a reference panel gave similar results in a GWAS for a polygenic trait but required some quality control in the identification of homozygous-by-descent segments. Thus, while using a reference panel remains the ideal approach, imputation in suitably large lcWGS datasets can provide sufficient accuracy given proper quality control.