Reference-based phasing using the Haplotype Reference Consortium panel.

Schematic of the Eagle2 core phasing algorithmGiven diploid genotypes from a target sample along with a haploid reference set of conditioning haplotypes, our algorithm proceeds in two steps. (a) We use the positional Burrows-Wheeler transform20 to generate a “hedge” of haplotype prefix trees rooted at markers spaced across the chromosome. These trees encode haplotype prefix frequencies, represented here with branch thicknesses. (b) We explore a small set of high-probability diplotypes (i.e., complementary pairs of phased haplotypes), estimating diplotype probabilities under a haplotype copying model by summing over possible recombination points. For each possible choice of recombination points, the HapHedge data structure allows rapid lookup of haplotype segment frequencies. (This illustration is meant to provide intuition for the overall approach; our optimized software implementation first “condenses” reference haplotypes based on the target genotypes. Details are provided in Supplementary Fig. 1 and the Supplementary Note.)

Running time and accuracy of reference-based phasing in UK Biobank benchmarksWe benchmarked Eagle2 and other available methods by phasing UK Biobank trio children using a reference panel generated from Nref = 15,000, 30,000, 50,000, or 100,000 other UK Biobank samples. (a) CPU time per target genome on a 2.27 GHz Intel Xeon L5640 processor. (We analyzed a total of 174,595 markers on chromosomes 1, 5, 10, 15, and 20, representing ≈25% of the genome, and scaled up running times by a factor of 4; see Supplementary Table 3 for details.) (b) Mean switch error rate over 70 European-ancestry trios; error bars, s.e.m. (c, d) CPU time and mean switch error rate as a function of the number of conditioning haplotypes used by SHAPEIT2 and Eagle2 (relative to the default values of K=100 and 10,000, respectively). Eagle1 does not have such a parameter, so we display its performance as a horizontal line. Numeric data and additional benchmarks varying the number of conditioning haplotypes used with Nref = 15,000, 50,000, and 100,000 are provided in Supplementary Table 2.

Accuracy of reference-based phasing in GERA benchmarksWe phased trio parents in each GERA sub-cohort using a reference panel generated from all other non-familial samples in the same sub-cohort. We ran each method with default parameter settings on all 22 autosomes and computed aggregate mean switch error rates; error bars, s.e.m. Standard errors for the European-ancestry sub-cohort are over 400 parent samples. Standard errors for the other three sub-cohorts are over 25 SNP blocks. Numeric data and additional benchmarks varying the number of conditioning haplotypes used by each method are provided in Supplementary Table 4.

Accuracy of reference-based phasing using the 1000 Genomes and HRC panelsWe phased 32 trio children from the 1000 Genomes CEU population using either the 1000 Genomes Phase 3 reference panel or the Haplotype Reference Consortium panel (excluding trios in either case). We analyzed chromosome 1, and to emulate a typical use case, we restricted the data to 31,853 markers (genotyped on 23 and Me chips). We plot mean switch error rates; error bars, s.e.m. over samples. Numeric data and additional benchmarks on other 1000 Genomes populations are provided in Supplementary Table 5.

Running time and accuracy of cohort-based phasing in the UK Biobank cohortWe benchmarked Eagle2 and other available phasing methods on N=5,000, 15,000 50,000, and 150,000 UK Biobank samples (including trio children and excluding trio parents). (a) Total wall clock time for genome-wide phasing on a 16-core 2.60 GHz Intel Xeon E5-2650 v2 processor. (We analyzed a total of 174,595 markers on chromosomes 1, 5, 10, 15, and 20, representing ≈25% of the genome, and scaled up running times by a factor of 4; see Supplementary Table 8 for per-chromosome data.) SHAPEIT2 was unable to complete the N=50,000 chr1 and chr5 analyses and was uanble to complete any of the N=150,000 analyses in 5 days, the run time limit for single compute jobs. (b) Mean switch error rate over 70 European-ancestry trios; error bars, s.e.m. Numeric data and additional benchmarks varying the number of conditioning haplotypes used by Eagle2 are provided in Supplementary Table 7.