the Eagle algorithm is in principle quite robust to error (Supplementary Table 15), but additional tuning will undoubtedly be necessary. Second, an imputation algorithm capable of rapidly and accurately imputing pre-phased target samples using very large imputation reference panels will be needed. Several efforts to develop such methods are currently underway: the Sanger Imputation Service (see URLs) is already using a new (unpublished) imputation algorithm based on the Positional Burrows-Wheeler Transformation (PBWT)37—which like Eagle applies fast string matching algorithms in favor of exact statistical modeling—and the Beagle v4.1 imputation software38 and the Minimac3 imputation software (unpublished but in use by the Michigan Imputation Server; see URLs) likewise aim to satisfy these requirements. Finally, the sequence data itself will need to be generated. However, very large scale sequencing projects are already underway: e.g., Genomics England plans to sequence 100,000 genomes by 2017 (see URLs).
