We next undertook a detailed comparison of phasing accuracy achieved by the two most accurate methods, Eagle and SHAPEIT2, when run on N≈150,000 samples. For this comparison, we analyzed ten 10,000-SNP regions (of median length 44Mb) comprising 16% of the genome (Supplementary Table 4). To overcome the high computational cost of SHAPEIT2 N≈150,000 analyses (Fig. 2a), we performed these benchmarks on the Lisa Genetic Cluster Computer (see URLs), which offered high-throughput parallel computation in batches of 16-core, 5-day jobs. Because SHAPEIT2 was unable to complete 10,000-SNP analyses within a single job, we split each 10,000-SNP analysis into three overlapping blocks of 3,667 SNPs (with an overlap of 500 SNPs); we ligated the results using hapfuse v1.6.2 (see URLs). In these benchmarks, we observed that Eagle achieved a slightly lower switch error rate than SHAPEIT2 run with default parameters (K=100 conditioning states12): 0.276% (0.008%) for Eagle vs. 0.306% (0.013%) for SHAPEIT2 (p=0.03, one-sided paired t-test) (Table 1). One caveat regarding these comparisons is that splitting and ligating the SHAPEIT2 analyses may have incurred a slight loss of accuracy12; while computational limitations prevented us from running SHAPEIT2 on full 10,000-SNP regions, performing the full analyses in single computations could improve accuracy.
