Using this simulation framework, iterated thousands of times over dozens of independent female and male samples, we find that for deletions spanning 3, 5 and 10 probes (corresponding to mean sizes of 5 kb, 8 kb and 17 kb), Birdseye identified with lod of 2 or greater 10%, 51% and 97.5% of the events, respectively; as expected, mean reported lod score also increased with deletion size (Supplementary Table 1 online). Breakpoints were typically determined to within a single probe of the simulated CNV, and fewer than one false positive is expected per genome at a lod of 2 or greater (Fig. 4d). We note that because the simulation removes local autocorrelation of noise, this may overestimate performance on actual data; higher lod cutoffs may be appropriate in different datasets. Nonetheless, this simulation indicates that the combination of the Affymetrix SNP 6.0 array and the Birdsuite seems highly sensitive for deletions of 10 kb or larger.
