proband29. In contrast, Eagle takes a “top-down” approach, first scanning all pairs of individuals for long IBD tracts and using them to phase long stretches of genome, and then applying only two iterations of approximate HMM decoding to correct errors and fill in unphased regions (Figure 1). Thus, at a high level, the key methodological contribution of Eagle's “top-down” approach is its use of LRP to greatly improve speed (by over an order of magnitude) by eliminating the need to slowly build phase accuracy over many HMM sampling iterations. This speedup is essential at large sample sizes: due to computational constraints, the production phasing of UK Biobank samples was not performed using the most accurate method available, SHAPEIT2; instead, a new (currently unreleased) method was developed, SHAPEIT3, which was reported to achieve a higher switch error rate of ≈0.4% (see URLs). At very large sample sizes to come, experience from Iceland indicates that HMM iterations may not be necessary at all13,15,25; instead, optimizing accuracy will require solving problems of a different nature, e.g., resolving conflicts in IBD information.
