Fast and accurate long-range phasing in a UK Biobank cohort

Abstract

Recent work has leveraged the extensive genotyping of the Icelandic population to perform long-range phasing (LRP), enabling accurate imputation and association analysis of rare variants in target samples typed on genotyping arrays. Here we develop a fast and accurate LRP method, Eagle, that extends this paradigm to populations with much smaller proportions of genotyped samples by harnessing long (>4-cM) identical-by-descent (IBD) tracts shared among distantly related individuals. We applied Eagle to N ≈ 150,000 samples (0.2% of the British population) from the UK Biobank, and we determined that it is 1-2 orders of magnitude faster than existing methods while achieving similar or better phasing accuracy (switch error rate ≈ 0.3%, corresponding to perfect phase in a majority of 10-Mb segments). We also observed that, when used within an imputation pipeline, Eagle prephasing improved downstream imputation accuracy in comparison to prephasing in batches using existing methods, as necessary to achieve comparable computational cost.

Figure 1. Eagle algorithm and example phase calls after each step

We show phase calls for ten trio children after each successive step of the Eagle algorithm (applied to phase the first 40cM of chromosome 10 in all N≈150,000 UK Biobank samples except trio parents). At all trio-phased sites, red and blue indicate whether the first Eagle-phased haplotype for each child matches the maternal or paternal haplotype. (a) After the first step, a sizable proportion of each genome is covered by long segments of near-perfect phase; these segments are the regions in which long IBD is available from several relatives. (b) The second step, which uses both long and short IBD, fixes most of the phase switch errors in the first step. (c,d) The subsequent approximate HMM iterations further reduce the error rate.

Figure 2. Computational cost and accuracy of phasing methods

Benchmarks of Eagle and existing phasing methods (all run with default options) on N≈15,000, 50,000, and 150,000 UK Biobank samples and M=5,824 SNPs on chromosome 10. Log-log plots of (a) run times and (b) memory consumption using up to 10 cores of a 2.27 GHz Intel Xeon L5640 processor and up to two weeks of computation. (c) Mean switch error rate over 70 European-ancestry trios; error bars, s.e.m. All methods except HAPI-UR supported multithreading. As the HAPI-UR documentation suggested merging results from three independent runs with different random seeds, we parallelized these runs across three cores. (For the N≈150,000 experiment, HAPI-UR encountered a failed assertion bug for some random seeds, so we needed to try six random seeds to find three working seeds. We did not count this extra work against HAPI-UR.) Numeric data are provided in Supplementary Table 1.