Performance of genotype imputations using data from the 1000 Genomes Project

Abstract

Genotype imputations based on 1000 Genomes (1KG) Project data have the advantage of imputing many more SNPs than imputations based on HapMap data. It also provides an opportunity to discover associations with relatively rare variants. Recent investigations are increasingly using 1KG data for genotype imputations, but only limited evaluations of the performance of this approach are available. In this paper, we empirically evaluated imputation performance using 1KG data by comparing imputation results to those using the HapMap Phase II data that have been widely used. We used three reference panels: the CEU panel consisting of 120 haplotypes from HapMap II and 1KG data (June 2010 release) and the EUR panel consisting of 566 haplotypes also from 1KG data (August 2010 release). We used Illumina 324,607 autosomal SNPs genotyped in 501 individuals of European ancestry. Our most important finding was that both 1KG reference panels provided much higher imputation yield than the HapMap II panel. There were more than twice as many successfully imputed SNPs as there were using the HapMap II panel (6.7 million vs. 2.5 million). Our second most important finding was that accuracy using both 1KG panels was high and almost identical to accuracy using the HapMap II panel. Furthermore, after removing SNPs with MACH Rsq <0.3, accuracy for both rare and low frequency SNPs was very high and almost identical to accuracy for common SNPs. We found that imputation using the 1KG-EUR panel had advantages in successfully imputing rare, low frequency and common variants. Our findings suggest that 1KG-based imputation can increase the opportunity to discover significant associations for SNPs across the allele frequency spectrum. Because the 1KG Project is still underway, we expect that later versions will provide even better imputation performance.

Fig. 1

Venn diagram showing 14,163,455 SNPs on chromosomes 1 through 22 across the three reference panels from HMII and 1KG Projects. For the overlap between 1KG-CEU and 1KG-EUR, the hg18 map positions of 1KG-CEU were converted into hg19 positions, using the liftOver program on the UCSC Genome Browser web site.

Fig. 2

Imputation yield (left) and accuracy (right) across the MAF spectrum for the 5% masked data using the three reference panels. Colored bars show yield and accuracy using filtered SNPs. Gray bars show total number of imputed SNPs and accuracy using all imputed SNPs. Online supplementary figure 1 shows the same information for all masked data.

Fig. 3

Imputation accuracy (dosage Rsq) values versus MACH Rsq values for the 5% masked data. Red solid circles are rare SNPs and black solid squares are low frequency SNPs. The black line indicates where MACH Rsq equals dosage Rsq. The magenta line is the regression line. The vertical dashed line is the filtering rule that we used. Imputation accuracy (table 4, fig. 2) was computed as the average of dosage Rsq values shown in the Y-axis. Colors refer to the online version only.