1-step versus 2-step imputation: a case study in German Black Pied cattle

Summary

Since the genotyping of cattle is usually performed using low or medium density SNP chips, the in silico imputation is a required step to generate genotypes on the whole-genome level using high density and whole-genome sequencing data.
Here, we investigate different imputation strategies using the recently available sequencing data from the 1000 bull genomes project including 30 sequenced DSN cattle (German Black Pied cattle, German: “Deutsches Schwarzbuntes Niederungsrind”) that we have contributed. Our investigation compares a 1-step versus a 2-step imputation approach using the imputation software tool Beagle. In the 1-step approach, 50k genotypes are directly imputed to the level of whole-genome sequencing data. The 2-step approach differs by first imputing 50k genotypes to 700k and subsequently from 700k to sequence level. Additionally, we investigate the imputation accuracy with respect to different reference population sizes and composition. These are 1) only DSN cattle, 2) DSN and Holstein cattle and 3) all Bos taurus cattle from the 1000 bull genomes project. The imputation accuracy was assessed as relative Manhattan distance in a leave-one-out cross validation using our 30 sequenced DSN cattle as targets for imputation.
Imputation with Beagle showed increased performance with increasing population sizes of the reference population, however a significant drop in imputation performance was observed when imputing using a smaller reference population consisting of breeds highly related to DSN compared to a larger reference population of unrelated breeds. Both size and composition play an important role in imputation accuracy. Furthermore, when using a small reference population in the first round (from 50k to 700k), we observed lower imputation accuracies of the 2-step approach compared to the 1-step approach. However, using a ‘big enough’ reference population in the first round restored imputation accuracies of the 2-step approach versus the much simpler 1-step approach. Our hypothesis is that when a limited reference population is available the 2-step approach leads to lower accuracy of imputation because imputation errors in the first round propagate to the second round of imputation.