In addition to reproducibility, we also sought to measure the accuracy of the breakpoints called by each platform. To perform this analysis, we used two data sets providing well-defined breakpoint information. First, we created a data set with nucleotide resolution breakpoints by combining data from capture and sequencing of CNV breakpoints43 with breakpoints collated44 from personal sequencing projects (Supplementary Fig. 12). The distance between the sequenced breakpoints and the CNV call coordinates in the present study was binned and plotted (Fig. 4a and Supplementary Methods). Only the more recent high-resolution arrays had enough CNV calls to yield meaningful results for this analysis (Supplementary Fig. 13). The data show that all platforms tend to underestimate the size of CNVs. This might be expected as the results reported for each algorithm correspond to the last probes within the variant showing evidence of a copy number difference. Arrays targeting known CNVs show the highest accuracy, as a result of their high probe density at these loci.
