larger than the reference size), we expect that reference-based phasing will achieve only marginal gains in accuracy over cohort-based phasing. For such cohorts, we expect that cohort-based phasing may remain the preferred option due to ease of execution—although reference sizes are growing, and if the end goal is HRC-based imputation, then we expect that HRC-based pre-phasing27 will be more convenient. Finally, we have not fully optimized Eagle2 for phasing sequenced samples28,29. While a preliminary benchmark against SHAPEIT2 suggests that Eagle2 achieves gains in speed and accuracy on sequence data comparable to its improvements on genotype data (Supplementary Table 10), more investigation will be needed to tune the method and benchmark it against other approaches (e.g., the recent SHAPEITR method30, which is available via the Oxford Statistics Phasing Server; see URLs). Despite these limitations, we expect that reference-based phasing using Eagle2 and the HRC panel will be a valuable resource providing free, fast, and accurate phasing to the scientific community.
