Lastly, we assessed the performance of Eagle2 when applied to cohort-based phasing, which we also implemented in our software. The Eagle2 cohort-based phasing algorithm starts by running the first two steps of Eagle113 to rapidly produce rough haplotype estimates and then refines these estimates using the Eagle2 core phasing algorithm (Online Methods). We benchmarked Eagle2, Eagle1, and SHAPEIT2 on subsets of the UK Biobank data set containing N = 5,000, 15,000, 50,000, or 150,000 samples (including trio children and excluding trio parents). We phased chromosomes 1, 5, 10, 15, and 20 as in our UK Biobank reference-based phasing benchmarks, and we allowed each computational job up to 5 days to complete. We observed that Eagle2 exhibited computational efficiency similar to Eagle1, achieving 5–6x speedups over SHAPEIT2 in the analyses SHAPEIT2 was able to complete (N=5,000 and N=15,000) (Fig. 5a and Supplementary Tables 7 and 8). Eagle2 also exhibited close-to-linear run time scaling across this sample size range, breaking even with Eagle1 at N≈30,000 and achieving faster running times for larger N.
